<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/">
    <channel>
        <title>Dwaar Blog</title>
        <link>https://docs.dwaar.ai/blog</link>
        <description>Dwaar Documentation Blog</description>
        <lastBuildDate>Thu, 12 Nov 2026 00:00:00 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>https://github.com/jpmonette/feed</generator>
        <language>en</language>
        <copyright>© 2026 Dwaar AI</copyright>
        <item>
            <title><![CDATA[When an Attack Costs $25, Compliance Is Not a Security Program]]></title>
            <link>https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program</link>
            <guid>https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program</guid>
            <pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[One attacker used AI agents to hit 105 retailers for $25.46 a company. A lab agent took a domain in eight minutes for $12. Cheap attacks change who gets targeted, and a passed audit does not change the answer.]]></description>
            <content:encoded><![CDATA[<p>Most smaller companies run on an unspoken assumption: we are not interesting enough to attack, so we will do what the audit asks and spend the rest on the product. That assumption was always shaky. It rested on the attacker's cost, because a skilled team's time had to be spent somewhere worth it. AI agents have just cut that cost to the price of lunch.</p>

































<table><thead><tr><th>At a glance</th><th></th></tr></thead><tbody><tr><td>Campaign</td><td>One attacker, three AI harnesses, open-weight and public models, 105 e-commerce sites in five days, per <a href="https://gambit.security/blog-posts/autonomous-ai-agents-online-retailers-25-a-company" target="_blank" rel="noopener noreferrer" class="">Gambit Security</a></td></tr><tr><td>Compromised</td><td>At least 27 companies, usually within a day</td></tr><tr><td>Cost to the attacker</td><td>$25.46 on average per completed target</td></tr><tr><td>Lab test</td><td>Full domain control in eight minutes for about $12, per <a href="https://www.silverfort.com/blog/5-cybersecurity-assumptions-ai-attacks-break/" target="_blank" rel="noopener noreferrer" class="">Silverfort</a></td></tr><tr><td>Price of initial access</td><td>Down from just over $1,400 to $429 on average, per Flashpoint</td></tr><tr><td>Reported</td><td><a href="https://www.darkreading.com/cybersecurity-operations/ai-tools-massive-attacks-little-cost" target="_blank" rel="noopener noreferrer" class="">Dark Reading</a>, 23 October 2026</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-25-bought">What $25 bought<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#what-25-bought" class="hash-link" aria-label="Direct link to What $25 bought" title="Direct link to What $25 bought" translate="no">​</a></h2>
<p>Over five days in September, a Chinese-speaking attacker pointed three AI harnesses at 105 e-commerce sites: the agentic system Hermes and two AI penetration testing frameworks, Strix and Cairn. At least 27 companies were compromised, usually within a day and often within hours. Agents stole at least 600,000 payment card records from two of them and installed card skimmers on the websites of five. The rest saw some degree of illicit access, and at least one was a Fortune 500 company.</p>
<p>The average bill was $25.46 per completed target.</p>
<p>"Work that would previously have required a coordinated team of skilled attackers can now be carried out by one person without much prior knowledge," Sa'ar Elias, co-founder and chief research officer at Gambit Security, told Dark Reading.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="eight-minutes-12">Eight minutes, $12<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#eight-minutes-12" class="hash-link" aria-label="Direct link to Eight minutes, $12" title="Direct link to Eight minutes, $12" translate="no">​</a></h2>
<p>Silverfort ran the controlled version. An orchestrating agent using publicly available versions of Claude Opus started from initial access and reached complete domain control in eight minutes. The total cost was about $12.</p>
<p>"The barrier to entry is falling," said Colin Ferris, threat hunting and incident response lead at Silverfort. "What it takes is a basic foothold, not exotic skill."</p>
<p>The market agrees. Flashpoint reports that the average price of initial access sold in underground markets fell from just over $1,400 to $429. Fewer companies paid ransoms, 28% in the first half of 2026, yet ransomware activity still grew by 45%.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="why-the-cost-curve-lands-on-smaller-companies">Why the cost curve lands on smaller companies<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#why-the-cost-curve-lands-on-smaller-companies" class="hash-link" aria-label="Direct link to Why the cost curve lands on smaller companies" title="Direct link to Why the cost curve lands on smaller companies" translate="no">​</a></h2>
<p>The data above says attacks got cheap. It does not break victims down by company size, so what follows is our reading of it, not a statistic.</p>
<p>When an attack took a skilled team days, attackers had to choose. Large targets justified the effort, and a small company was protected mostly by being not worth the time. At $25 a target, that filter is gone. An agent does not get bored of small websites, and 105 retailers in five days is what that looks like. Albert Ziegler, CISO at XBOW, put it plainly: every part of a company attached to the internet will be attacked, and the attacks in this campaign "work because the sites haven't been hardened enough."</p>
<p>Large enterprises felt this too, and the Fortune 500 victim proves it. But they have more to absorb the hit with: a security team, someone on call, budget for incident response. A smaller company often has a compliance checklist and an annual audit. When obscurity was doing the real work, that gap stayed hidden. Cheap attacks expose it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="a-passed-audit-is-not-hardening">A passed audit is not hardening<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#a-passed-audit-is-not-hardening" class="hash-link" aria-label="Direct link to A passed audit is not hardening" title="Direct link to A passed audit is not hardening" translate="no">​</a></h2>
<p>Compliance asks whether a control exists and whether you can show evidence for it on the day of the audit. Attackers ask whether it works today. Those questions overlap less than most teams think.</p>
<p>Look at what Silverfort says AI agents exploit: overprivileged accounts, reused credentials and unmanaged service accounts. None of these is exotic. All of them can survive an audit, because an access review done once a quarter is a passing control even when a service account has had domain admin rights since last year. "AI didn't create these weaknesses; it just finds them faster," Ferris said.</p>
<p>The encouraging part is in the same test. When Silverfort tightened controls, especially authentication and access, the same agent that took the domain in minutes hit 21 dead ends and gave up. The attacker's cost went up the moment the basics held. That is the lever a smaller company can actually pull.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="spend-on-what-keeps-the-business-running">Spend on what keeps the business running<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#spend-on-what-keeps-the-business-running" class="hash-link" aria-label="Direct link to Spend on what keeps the business running" title="Direct link to Spend on what keeps the business running" translate="no">​</a></h2>
<p>Elias's advice is economic, and it scales down well. Defending against every attempt is unsustainable when each attempt costs the attacker so little. Instead, work out which business processes matter most and protect those from disruption. "Spending should increasingly be measured by how effectively it protects business operations," he said, "rather than simply how many threats or tools a security program can manage."</p>
<p>For a smaller team, that turns into a short list:</p>
<ol>
<li class=""><strong>Name your minimum viable business.</strong> Checkout, payments, customer data, the admin panel. If the agent in this campaign found yours, what would stop it?</li>
<li class=""><strong>Fix the paths agents take first.</strong> Remove standing admin rights, retire reused and shared credentials, and give every service account an owner. Turn on strong authentication everywhere an attacker could start.</li>
<li class=""><strong>Check continuously, not at audit time.</strong> An attacker running agents tests you every day. A quarterly review leaves the other 89 days unchecked. Use automated checks that confirm your controls still hold, and keep audit evidence as what those checks produce, not as a separate project.</li>
<li class=""><strong>Treat every internet-facing asset as in scope.</strong> Ziegler's warning applies here: protecting a couple of crown jewel sites is not enough. The forgotten storefront or staging app is exactly what a $25 scan finds.</li>
</ol>
<p><a class="" href="https://docs.dwaar.ai/blog/why-remediation-has-to-move-at-agentic-speed">Remediation has to move at agentic speed</a>, because discovery already does. And <a class="" href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities">counting vulnerabilities will not tell you</a> whether the account an agent would use next is locked down.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-this-next">Try this next<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#try-this-next" class="hash-link" aria-label="Direct link to Try this next" title="Direct link to Try this next" translate="no">​</a></h2>
<p>Pick one service account in your most important system and answer three questions: who owns it, what can it reach, and when did anyone last check? If any answer is "we would have to look", you have found the eight minute path. Close that one before the next audit asks about it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/when-an-attack-costs-25-dollars-compliance-is-not-a-security-program#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://www.darkreading.com/cybersecurity-operations/ai-tools-massive-attacks-little-cost" target="_blank" rel="noopener noreferrer" class="">AI Tools Are Generating Massive Attacks for Little Cost</a>, Dark Reading, 23 October 2026</li>
<li class=""><a href="https://gambit.security/blog-posts/autonomous-ai-agents-online-retailers-25-a-company" target="_blank" rel="noopener noreferrer" class="">Gambit Security's research on the retail campaign</a>, Gambit Security</li>
<li class=""><a href="https://www.silverfort.com/blog/5-cybersecurity-assumptions-ai-attacks-break/" target="_blank" rel="noopener noreferrer" class="">Silverfort's research on the domain takeover test</a>, Silverfort</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>AI agents</category>
            <category>Compliance</category>
        </item>
        <item>
            <title><![CDATA[Trusting a Leading Agentic Harness Does Not Guarantee the Security of Your Application]]></title>
            <link>https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security</link>
            <guid>https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security</guid>
            <pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Claude Code's "don't ask again" file put 1,379 secrets into 1,122 public GitHub copies. The harness did try to protect them. The lesson is that the control you rely on has to live outside the agent, and across every team.]]></description>
            <content:encoded><![CDATA[<p>Picking the most capable coding agent on the market is a reasonable decision. Assuming that decision also settles your security posture is not. This week's research on Claude Code shows why. One click on a permission prompt can write a production password into a file, and that file can end up in a public repository. The harness had a safeguard, and the safeguard ran on one laptop.</p>

































<table><thead><tr><th>At a glance</th><th></th></tr></thead><tbody><tr><td>The file</td><td><code>.claude/settings.local.json</code>, written when a user picks <strong>Yes, and don't ask again</strong></td></tr><tr><td>Public copies found</td><td>104,185 in 103,617 repositories, per <a href="https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again" target="_blank" rel="noopener noreferrer" class="">Apiiro</a></td></tr><tr><td>Copies holding a secret</td><td>1,122, about 1.1%, holding 1,379 different secrets</td></tr><tr><td>Total across agent files, history and chat logs</td><td>2,539 secrets in 1,994 repositories</td></tr><tr><td>Missed by both gitleaks and TruffleHog</td><td>215 of the 1,122 files</td></tr><tr><td>Reported</td><td>Apiiro on 6 November 2026, covered by <a href="https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html" target="_blank" rel="noopener noreferrer" class="">The Hacker News</a> on 7 November</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="how-a-click-becomes-a-leaked-password">How a click becomes a leaked password<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#how-a-click-becomes-a-leaked-password" class="hash-link" aria-label="Direct link to How a click becomes a leaked password" title="Direct link to How a click becomes a leaked password" translate="no">​</a></h2>
<p>By default, Claude Code asks before running most shell commands. Choose <strong>Yes, and don't ask again</strong> and it saves an allow rule in <code>.claude/settings.local.json</code> inside the project, per its <a href="https://code.claude.com/docs/en/permissions" target="_blank" rel="noopener noreferrer" class="">permissions documentation</a>. Matching commands then run without asking in later sessions.</p>
<p>The rule covers whatever the prompt offered. Sometimes that is a pattern such as <code>npm test *</code>. Sometimes it is one exact command, and then any password or token typed into that command is saved with it, in plain text. The researchers found saved commands with the secret still inside in public files, and reproduced the behavior with Opus 5.5 and Fable 5.1, the latest models.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-harness-did-try">The harness did try<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#the-harness-did-try" class="hash-link" aria-label="Direct link to The harness did try" title="Direct link to The harness did try" translate="no">​</a></h2>
<p>This is the part worth sitting with. Claude Code does not ignore the risk. The first time it writes the file in a git repository, it adds the file to the user's global git excludes, per its <a href="https://code.claude.com/docs/en/settings" target="_blank" rel="noopener noreferrer" class="">settings documentation</a>. On that computer, git then ignores the file in every repository.</p>
<p>That protection stays in one user's git settings on one machine. It does nothing for a file that was already committed. Claude Code never adds the file to the project's own <code>.gitignore</code>, so the protection does not travel with the project to a teammate, a CI runner or a fresh clone.</p>
<p>The agent also helped the file get committed. In 539 of the 1,105 repositories the researchers traced, a commit listing Claude as a co-author changed the file. Apiiro reads this as Claude often committing the file itself.</p>
<p>So a well funded vendor shipped a sensible mitigation, and it still leaked 1,379 secrets. The issue is not one careless vendor. A safeguard that lives inside one tool, on one machine, configured by one person, cannot be an organization's control.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-leaked-and-why-it-matters">What leaked, and why it matters<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#what-leaked-and-why-it-matters" class="hash-link" aria-label="Direct link to What leaked, and why it matters" title="Direct link to What leaked, and why it matters" translate="no">​</a></h2>
<p>More than a third of the secrets open a database directly, including 379 database passwords. The rest include 43 SSH logins for public servers, 42 tokens for GitHub, GitLab and npm, and 23 AWS keys. The owners include 63 organizations, mostly software companies. The strongest cases include a database vendor's admin-level cloud key and the owner password of a medical records database.</p>
<p>Three numbers say this is not a cleanup job you can finish by deleting a file:</p>





















<table><thead><tr><th>Where the secret still lives</th><th>What Apiiro found</th></tr></thead><tbody><tr><td>Default branch</td><td>Every file with a secret was still there as of 2 November. In half of those repositories, it had been there for at least 217 days.</td></tr><tr><td>Git history</td><td>734 more secrets across 509 repositories exist only in older versions of the file.</td></tr><tr><td>Forks</td><td>211 forks of 74 repositories carry the same secrets as the originals.</td></tr></tbody></table>
<p>Apiiro did not test whether these secrets still work. One company it notified confirmed its key had been live.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="your-scanner-probably-missed-it">Your scanner probably missed it<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#your-scanner-probably-missed-it" class="hash-link" aria-label="Direct link to Your scanner probably missed it" title="Direct link to Your scanner probably missed it" translate="no">​</a></h2>
<p>Of the 1,122 files with a secret, gitleaks flagged 564 and TruffleHog flagged 643, run without their verification step. Neither flagged 215. Most misses are passwords typed inside commands. Scanners look for known key formats, and a password after <code>PGPASSWORD=</code>, <code>sshpass -p</code> or <code>mysql -p</code> looks like an ordinary word in an ordinary command line.</p>
<p>That gap is not new territory. <a class="" href="https://docs.dwaar.ai/blog/secrets-live-in-more-places-than-your-repository">Secrets live in more places than your repository</a>, and an agent's config folder is now one of them. Apiiro notes the file can also travel inside an npm package, a Docker image, a zip archive or a synced folder.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="it-was-reported-and-closed">It was reported, and closed<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#it-was-reported-and-closed" class="hash-link" aria-label="Direct link to It was reported, and closed" title="Direct link to It was reported, and closed" translate="no">​</a></h2>
<p>A <a href="https://github.com/anthropics/claude-code/issues/5544" target="_blank" rel="noopener noreferrer" class="">bug report filed on 11 August 2025</a> said approved commands with API keys and tokens were stored word for word in <code>settings.local.json</code>. An automated bot closed it as "not planned" on 3 January, after 60 days without activity. Apiiro says the four related reports it cites were all closed by the same bot, and none got a reply from Anthropic.</p>
<p>The point is not to single out Anthropic. Of 16 coding agents whose documentation Apiiro compared, Claude Code is the only one confirmed to save the full approved command in a project file by default. The other 15 have their own defaults, and your teams probably run more than one of them. You do not control any vendor's roadmap or triage queue. You control what you check.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="why-the-fix-cannot-be-another-agent">Why the fix cannot be another agent<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#why-the-fix-cannot-be-another-agent" class="hash-link" aria-label="Direct link to Why the fix cannot be another agent" title="Direct link to Why the fix cannot be another agent" translate="no">​</a></h2>
<p>The tempting answer is to ask the agent to be more careful, or to add an AI reviewer. Both put the control back inside a system that makes a judgment call each time, on one developer's machine, in one session. That is the same shape that failed here.</p>
<p>What this needs is boring and deterministic:</p>
<ul>
<li class=""><strong>The same rule in every repository.</strong> Whether a team uses Claude Code, Cursor or Gemini, the check for agent config files with secrets in them runs the same way, every time, without anyone opting in.</li>
<li class=""><strong>History and forks, not just HEAD.</strong> A rule that only reads the default branch misses 734 secrets in this study alone.</li>
<li class=""><strong>Organization templates, not personal settings.</strong> The harness protected one laptop. Best practice has to sit in the repository and in the pipeline, where every clone and every CI run inherits it.</li>
<li class=""><strong>Evidence for every team.</strong> A security lead should be able to answer "which projects carry <code>.claude/settings.local.json</code> in history" across the whole organization, not by asking each team.</li>
</ul>
<p><a class="" href="https://docs.dwaar.ai/blog/the-agent-harness-is-where-ai-starts-acting-security-has-to-start-there-too">Security has to start at the agent harness</a>, because that is where the agent acts. But it cannot end there, because the harness is software someone else ships, with defaults someone else chose. The check that confirms your best practices hold has to sit outside the agent, and it has to be indifferent to which agent each team picked. That begins with <a class="" href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run">knowing which AI agents you run</a>.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-do-this-week">What to do this week<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#what-to-do-this-week" class="hash-link" aria-label="Direct link to What to do this week" title="Direct link to What to do this week" translate="no">​</a></h2>
<p>Apiiro's guidance, in the order that limits damage:</p>
<ol>
<li class=""><strong>Find every copy.</strong> Search your GitHub organization for <code>.claude/settings.local.json</code> and <code>.claude/settings.json</code>, including git history where the files are gone.</li>
<li class=""><strong>Rotate first.</strong> Change every password, token and key the files have ever held, before you clean up. Deleting the file leaves history and forks untouched.</li>
<li class=""><strong>Stop tracking the local file.</strong> Run <code>git rm --cached .claude/settings.local.json</code> and add it to the repository's own <code>.gitignore</code>. Add it to your organization's <code>.gitignore</code> templates too.</li>
<li class=""><strong>Clean the file on every laptop.</strong> Delete every rule that contains a secret, even where the file was never committed.</li>
<li class=""><strong>Keep secrets out of commands.</strong> Reference them as variables from a secrets manager, such as <code>$DB_URL</code>, so the saved rule holds the name and not the value. Approve patterns such as <code>Bash(npm run:*)</code> instead of single commands.</li>
<li class=""><strong>Scan agent folders like <code>.env</code> files.</strong> Treat <code>.claude/</code>, <code>.cursor/</code>, <code>.gemini/</code>, <code>.specstory/</code> and agent memory files as holding secrets, in pre-commit and CI scanning, including history. Make sure the scanner catches passwords inside commands, not only known key formats. Check credentials paired with a public server first, because those give access.</li>
</ol>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-this-next">Try this next<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#try-this-next" class="hash-link" aria-label="Direct link to Try this next" title="Direct link to Try this next" translate="no">​</a></h2>
<p>Pick the three busiest repositories in your organization and look for <code>.claude/settings.local.json</code> in the full git history, not just the current tree. Then ask how you would know if a fourth team committed one tomorrow. If the honest answer is "we would find out when someone noticed", the agent you trust is not the gap. The missing control around it is.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/trusting-a-leading-agentic-harness-does-not-guarantee-application-security#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again" target="_blank" rel="noopener noreferrer" class="">Research write-up on secrets in "don't ask again" files</a>, Apiiro, 6 November 2026</li>
<li class=""><a href="https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html" target="_blank" rel="noopener noreferrer" class="">Claude Code's "Don't Ask Again" File Exposes Passwords and Keys in 1,122 Public Copies on GitHub</a>, The Hacker News, 7 November 2026</li>
<li class=""><a href="https://code.claude.com/docs/en/permissions" target="_blank" rel="noopener noreferrer" class="">Claude Code permissions</a>, Anthropic</li>
<li class=""><a href="https://code.claude.com/docs/en/settings" target="_blank" rel="noopener noreferrer" class="">Claude Code settings</a>, Anthropic</li>
<li class=""><a href="https://github.com/anthropics/claude-code/issues/5544" target="_blank" rel="noopener noreferrer" class="">anthropics/claude-code issue 5544</a>, bug report filed 11 August 2025</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>AI agents</category>
            <category>Identity</category>
        </item>
        <item>
            <title><![CDATA[Reachability Analysis: Why Most Dependency CVEs Don't Matter, and How to Prove It]]></title>
            <link>https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter</link>
            <guid>https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter</guid>
            <pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Most of the CVEs a dependency scanner reports are in code the application never executes. How reachability analysis proves it, what the research measures, and where static call graphs fail.]]></description>
            <content:encoded><![CDATA[<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-noise-problem">The noise problem<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#the-noise-problem" class="hash-link" aria-label="Direct link to The noise problem" title="Direct link to The noise problem" translate="no">​</a></h2>
<p>Most of the CVEs a dependency scanner reports are in code the application never executes. That is not an opinion about tooling; it is what every large-scale measurement of the problem has found, across ecosystems and years of studies.</p>
<p>Start with how much code an application actually pulls in. A 2019 study of the npm registry (676,539 packages, 5.4 million versions) found that installing an average package implicitly trusts 79 third-party packages and 39 maintainers, and that up to 40% of all packages depend on code with at least one publicly known vulnerability (<a href="https://usenix.org/system/files/sec19-zimmermann.pdf" target="_blank" rel="noopener noreferrer" class="">Zimmermann et al., USENIX Security 2019</a>). For Java, a 2023 study of 3 million Maven packages against 1,300 security reports found that of 1.1 million projects flagged as vulnerable, only 31% had the vulnerability in a <em>direct</em> dependency; the rest inherited it transitively (<a href="https://arxiv.org/abs/2301.07972" target="_blank" rel="noopener noreferrer" class="">Mir, Keshani and Proksch, SANER 2023</a>).</p>
<p>Now ask how much of that flagged code is ever called. The same Maven study built whole-program call graphs for the affected projects and found that only 1.2% of the 1.1 million transitively affected projects actually invoke the vulnerable code. A 2026 study of 2,414 open-source repositories across Python, Rust, Ruby and PHP went further: it generated accurate SBOMs from lock files, ran two mainstream scanners, and manually verified every reported vulnerability in a 16-repository sample. Of 113 reported vulnerabilities, 9 were real — a 92.0% false-positive rate — and the authors identified unreachable code as the primary cause (<a href="https://arxiv.org/pdf/2511.20313" target="_blank" rel="noopener noreferrer" class="">Zhou, Dacier and Konstantinou, CODASPY 2026</a>).</p>
<p>Two things to hold onto, then. In the Maven study roughly two thirds of flagged projects had the vulnerability only through a transitive dependency, so direct-only analysis leaves much of the problem unexamined. And when the analysis is done at the level of functions instead of packages, most of those flags turn out to be noise, by amounts that vary with the ecosystem, the method and what is being counted. Reachability analysis is the technique that measures this instead of guessing at it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-reachability-means-precisely">What reachability means, precisely<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#what-reachability-means-precisely" class="hash-link" aria-label="Direct link to What reachability means, precisely" title="Direct link to What reachability means, precisely" translate="no">​</a></h2>
<p>A vulnerability is <em>reachable</em> if there exists a path in the program's call graph from an entry point to a function the advisory identifies as vulnerable. Three terms in that sentence carry all the weight.</p>
<p><strong>Entry points</strong> are where execution can begin: <code>main</code>, package initialisers, exported handlers, framework-registered routes, test functions if you choose to count them. Miss an entry point and everything it alone reaches is wrongly classified unreachable. For a Go binary the set is well defined; for a Spring application half the entry points are discovered by the framework at runtime through annotations and the analysis must model that or admit it cannot.</p>
<p><strong>The call graph</strong> is a directed graph whose nodes are functions and whose edges are possible calls. The word <em>possible</em> is the whole subject. A direct call <code>f()</code> is one edge. A call through an interface, a virtual method, a function pointer or a dynamically named property may have one target or fifty, and the analysis has to decide which to include.</p>
<p><strong>Vulnerable symbols</strong> are the functions, methods or properties an advisory actually affects. Advisories name package version ranges; the symbol list has to come from somewhere else, which is the subject of a later section.</p>
<p>Two properties describe how good a call graph is, and they pull in opposite directions:</p>
<ul>
<li class=""><strong>Soundness</strong> (also called recall): every edge that can occur at runtime is in the graph. A sound graph never calls something unreachable when it is reachable.</li>
<li class=""><strong>Precision</strong>: edges in the graph can actually occur. An imprecise graph has edges to targets that never execute, so it calls things reachable that are not.</li>
</ul>
<p>For vulnerability triage the asymmetry is clear: a false <em>unreachable</em> lets a real CVE be waived, a false <em>reachable</em> costs a developer some minutes. So a reachability analysis for security must lean toward soundness, and the honest output has three values, not two:</p>

























<table><thead><tr><th>Verdict</th><th>Meaning</th><th>What the analysis must have established</th></tr></thead><tbody><tr><td>Reachable</td><td>A path exists from an entry point to a vulnerable symbol</td><td>At least one witness path, shown to the user</td></tr><tr><td>Unreachable</td><td>No such path exists in a graph the analysis trusts</td><td>Complete entry points, a sound graph for this package, no unresolved calls into it</td></tr><tr><td>Unknown</td><td>The analysis could not decide</td><td>A named reason: no symbol data, unresolved dynamic call, unsupported construct</td></tr></tbody></table>
<p>Any tool that returns only reachable or unreachable is hiding the third row inside one of the first two. Which one it hides it in tells you whether the tool was built for precision or for safety.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="how-a-call-graph-is-built">How a call graph is built<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#how-a-call-graph-is-built" class="hash-link" aria-label="Direct link to How a call graph is built" title="Direct link to How a call graph is built" translate="no">​</a></h2>
<p>The hard part of a call graph is not the direct calls; it is deciding what a polymorphic call site can invoke. Four classical algorithms answer that question with increasing precision and increasing cost, and every production reachability tool is one of them, or a hybrid.</p>








































<table><thead><tr><th>Algorithm</th><th>How it resolves <code>obj.method()</code></th><th>Precision</th><th>Cost</th><th>Where it is used</th></tr></thead><tbody><tr><td>Class Hierarchy Analysis (CHA)</td><td>Every subtype of the declared type that overrides <code>method</code></td><td>Low: a call on <code>Reader</code> targets every <code>Reader</code> in the program</td><td>Linear in program size; seconds on millions of lines</td><td>First pass everywhere; the over-approximate baseline</td></tr><tr><td>Rapid Type Analysis (RTA)</td><td>CHA, restricted to types that are instantiated somewhere reachable</td><td>Medium</td><td>Linear, with one fixpoint</td><td>Default in several JVM toolchains</td></tr><tr><td>Variable Type Analysis (VTA)</td><td>Propagates the set of types each variable can hold through assignments</td><td>Medium-high</td><td>Near-linear; one pass over an assignment graph</td><td>Go's vulnerability checker uses CHA then VTA</td></tr><tr><td>Points-to analysis (Andersen, Steensgaard, k-CFA)</td><td>Tracks which allocation sites each pointer can refer to, optionally per calling context</td><td>High</td><td>Cubic in the worst case; minutes to hours on large programs</td><td>Research tools, taint analysis, high-value targets</td></tr></tbody></table>
<p>The ordering matters because it is a soundness contract: CHA must over-approximate RTA, which must over-approximate VTA. A 2026 metamorphic study of Java frameworks found that this contract is routinely broken in practice — a more precise setting sometimes <em>adds</em> edges, and different frameworks disagree about what the ground truth even is (<a href="https://arxiv.org/pdf/2604.00885" target="_blank" rel="noopener noreferrer" class="">Zhong, Wold and Windmann, 2026</a>). The practical lesson is that the algorithm name on the box tells you less than a benchmark does.</p>
<p>For dynamically typed languages the type hierarchy does not exist, so the equivalent of CHA is name matching (any property called <code>parse</code> may call any function called <code>parse</code>), and anything better requires tracking values. The picture below shows what the graph is for.</p>
<p><img decoding="async" loading="lazy" alt="Only one of two vulnerable functions is on a path from the entry point" src="data:image/svg+xml;base64,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" width="760" height="490" class="img_ev3q"></p>
<p><em>illustrative call graph · two advisories, one witness path</em></p>
<p>Two advisories, one application. The package-level scanner reports both. The call graph shows that <code>main</code> reaches <code>OrderService.submit</code>, which calls the direct dependency's <code>http.Client.Do</code>, which reaches <code>h2.parseHeaders</code> inside a transitive dependency: CVE A is reachable, with that chain as evidence. <code>xml.expandEntity</code> is also present in the build and also vulnerable, but the only path to it starts at <code>ReportUtil.format</code>, which nothing calls: CVE B is unreachable. The grey nodes are what an SBOM counts; the blue path is what an attacker can use.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="modular-analysis-summarise-once-reuse-everywhere">Modular analysis: summarise once, reuse everywhere<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#modular-analysis-summarise-once-reuse-everywhere" class="hash-link" aria-label="Direct link to Modular analysis: summarise once, reuse everywhere" title="Direct link to Modular analysis: summarise once, reuse everywhere" translate="no">​</a></h2>
<p>A whole-program call graph of an application plus all of its dependencies is expensive, and almost all of that expense is spent on code that is identical across every application using the same package version. The scaling trick that makes reachability practical is to analyse each package once, in isolation, and keep only a <em>summary</em>: for every exported symbol, which vulnerable symbols it can reach, through its own transitive dependencies included.</p>
<p>The idea was formalised for Node.js by Nielsen, Torp and Møller in 2021. Their modular call-graph construction analysed each npm module separately, composed the results per application, and on a benchmark of real applications reduced the number of false positives by 81% compared with the package-level audit tool, with zero false negatives on the ground truth (<a href="https://www.doi.org/10.1145/3460319.3464836" target="_blank" rel="noopener noreferrer" class="">ISSTA 2021</a>). Analysis time also dropped substantially when modular results were reused across applications, because the per-package work is amortised.</p>
<p>What the application-side analysis then needs is small:</p>
<ol>
<li class="">The application's own call graph, from its entry points.</li>
<li class="">The set of (package, exported symbol) pairs the application calls directly.</li>
<li class="">For each direct dependency version, the precomputed summary.</li>
</ol>
<p>Reachability becomes a join: a vulnerability is reachable if some exported symbol the application calls has the vulnerable symbol in its reach set. The witness path is the application's own path to the call site, concatenated with the path stored in the summary. Transitive dependencies are never opened on the application side at all.</p>
<p>The approach has a cost in precision that is worth stating. A summary is context-insensitive: it says <code>Client.Do</code> <em>can</em> reach <code>parseHeaders</code>, not that it does so under this application's arguments and configuration. That is a source of false <em>reachable</em> verdicts, which is the safe direction. A VEX benchmark published in 2026 on Go projects shows the effect: about 70% of the false positives that reachability tools still produced were labelled code not reachable. The authors explain this by two things: a vulnerable function appearing in the call graph does not establish that the path is feasible under real branch conditions, and call graphs that are both sound and precise are hard to build (<a href="https://arxiv.org/pdf/2609.08040" target="_blank" rel="noopener noreferrer" class="">VEX-Bench, 2026</a>). Closing that gap requires data-flow or dynamic analysis per application, which is where the cost curve turns steep again.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-vulnerable-symbol-problem">The vulnerable-symbol problem<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#the-vulnerable-symbol-problem" class="hash-link" aria-label="Direct link to The vulnerable-symbol problem" title="Direct link to The vulnerable-symbol problem" translate="no">​</a></h2>
<p>A call graph answers "can I reach function F?" Advisories almost never say which F. A CVE record names a package and a version range; the function, method or property that actually carries the flaw is in the prose, the fix commit, or nowhere. Without that mapping the most precise call graph in the world has nothing to look for.</p>
<p>The Go ecosystem is the exception that proves the point. Its vulnerability database publishes structured affected-symbol lists per advisory, which is why fully automated reachability has existed there for years. The 2026 SBOM study above built a proof-of-concept on exactly that data and, across 12 widely used Go projects, pruned 74.4% of reported vulnerabilities (121 reported, 31 with the affected symbol confirmed present in the compiled binary). The authors confirmed with maintainers of major scanning tools that the absence of function-level metadata in mainstream advisory databases is a known, critical bottleneck, and they call for a standard <code>affected_symbols</code> field (<a href="https://arxiv.org/pdf/2511.20313" target="_blank" rel="noopener noreferrer" class="">Zhou et al., 2026</a>).</p>
<p>Until that field exists, symbols are derived. The standard method comes from the Java work of Ponta, Plate and Sabetta: identify the fix commit for the advisory, diff the vulnerable and fixed versions, and treat the changed program constructs, such as methods, as the vulnerable set. Their code-centric approach, which combined that derivation with static and dynamic reachability, was evaluated on more than 250,000 scans of about 500 applications in an industrial setting (<a href="https://arxiv.org/pdf/1806.05893" target="_blank" rel="noopener noreferrer" class="">ICSME 2018</a>). The same group released a manually curated dataset of fix commits to make the derivation reproducible (<a href="https://arxiv.org/abs/1902.02595" target="_blank" rel="noopener noreferrer" class="">MSR 2019</a>).</p>
<p>Fix-commit diffing is a heuristic, and its failure modes are worth knowing:</p>
<ul>
<li class="">The fix touches more than the flaw: refactoring, tests and helper changes inflate the symbol set, which is safe but noisy.</li>
<li class="">The fix touches less than the flaw: a one-line guard added in a caller leaves the real vulnerable function unchanged, so the set misses it. This is the unsafe direction.</li>
<li class="">The vulnerability is in initialisation, configuration or data rather than in a function: prototype pollution, unsafe defaults, a bundled native library. The honest symbol for these is "the module itself", which collapses reachability back to "is it imported".</li>
<li class="">The advisory has no public fix, or the fix is in a fork.</li>
</ul>
<p>A workable pipeline therefore has three tiers of confidence — authoritative (the ecosystem publishes symbols), reviewed (a human checked a derived set against the diff), derived (automated only) — and a rule that a <em>derived</em> set may produce a reachable verdict but never an unreachable one. The number of advisories that matter is also far smaller than the number that exist: the symbol work only has to be done for advisories that match packages real applications actually use, which in practice concentrates on a few hundred popular libraries per ecosystem.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-the-numbers-say">What the numbers say<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#what-the-numbers-say" class="hash-link" aria-label="Direct link to What the numbers say" title="Direct link to What the numbers say" translate="no">​</a></h2>
<p>The effect is large, consistent across projects, and bounded: reachability removes most of the noise, not all of it, and the residue is where the real work is.</p>
<p><img decoding="async" loading="lazy" alt="Symbol-level reachability pruned 74% of reported vulnerabilities across 12 Go projects" src="data:image/svg+xml;base64,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" width="760" height="400" class="img_ev3q"></p>
<p><em>Zhou, Dacier and Konstantinou, CODASPY 2026, Table 9 · 12 Go projects, 121 reported vulnerabilities; one project with zero reports omitted</em></p>
<p>The spread is the interesting part. Prometheus loses 90% of its reported vulnerabilities to reachability; Caddy loses only half. I read that spread as architecture rather than analysis quality: a web server that exposes much of its dependency surface at runtime has more live paths than a metrics collector that uses a narrow slice of each library. Reachability is measuring the application, not the scanner.</p>
<p>Across the literature the headline figures line up, even though they measure slightly different things:</p>















































<table><thead><tr><th>Study</th><th>Ecosystem</th><th>What was measured</th><th>Result</th></tr></thead><tbody><tr><td><a href="https://arxiv.org/abs/2301.07972" target="_blank" rel="noopener noreferrer" class="">Mir et al., SANER 2023</a></td><td>Maven, 3M packages</td><td>Transitively affected projects that call vulnerable code</td><td>1.2%</td></tr><tr><td><a href="https://www.doi.org/10.1145/3460319.3464836" target="_blank" rel="noopener noreferrer" class="">Nielsen et al., ISSTA 2021</a></td><td>npm</td><td>False positives removed vs package-level audit; false negatives</td><td>81% removed; 0 missed</td></tr><tr><td><a href="https://arxiv.org/pdf/2511.20313" target="_blank" rel="noopener noreferrer" class="">Zhou et al., CODASPY 2026</a></td><td>Python, Rust, Ruby, PHP; 16 repos, manual</td><td>Reported vulnerabilities confirmed real</td><td>9 of 113 (92.0% false positive)</td></tr><tr><td>Same study</td><td>Same sample, function-call check</td><td>Reports removed by checking whether the vulnerable function is referenced</td><td>113 → 42 (the paper reports 61.9%)</td></tr><tr><td>Same study</td><td>Go, 12 projects, automated</td><td>Reports removed by symbol-level analysis of the compiled binary</td><td>74.4% (121 → 31)</td></tr><tr><td><a href="https://arxiv.org/pdf/2609.08040" target="_blank" rel="noopener noreferrer" class="">VEX-Bench, 2026</a></td><td>Go, 30 cases</td><td>False positives remaining after reachability that were labelled code not reachable</td><td>about 70%</td></tr></tbody></table>
<p>Read together: package-level matching is wrong most of the time. Function-level reachability removes a large share of that error, from 71 of 104 false reports in the manually verified sample to 81% in the npm benchmark, but it is not perfect in either direction. In that same sample 33 of the 42 reports left after the function check were still false, and VEX-Bench recorded one reachability tool missing 1 of 7 truly affected Go cases. What remains includes paths that exist in the graph but cannot execute, which only data-flow or runtime evidence can settle.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-static-reachability-fails">Where static reachability fails<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#where-static-reachability-fails" class="hash-link" aria-label="Direct link to Where static reachability fails" title="Direct link to Where static reachability fails" translate="no">​</a></h2>
<p>Every static call graph is unsound somewhere, and for security the unsound places are the ones that matter, because they are where a real path can hide. The failure modes are well catalogued; what distinguishes a trustworthy tool is whether it reports them as Unknown or silently as Unreachable.</p>
<p><strong>Dynamic dispatch the analysis cannot resolve.</strong> Reflection in Java (<code>Class.forName</code>, <code>Method.invoke</code>), <code>getattr</code> and <code>importlib</code> in Python, computed property access and <code>eval</code> in JavaScript, function pointers cast through <code>void*</code> in C. A 2018 test-suite study of Java frameworks found that the two most widely used analysis frameworks each missed expected edges, that standard language features were well supported while support for core APIs varied significantly, and that support for lambdas and reflection differed between them (<a href="https://2018.ecoop.org/event/soap-2018-papers-systematic-evaluation-of-the-soundiness-of-call-graph-construction-algorithms-for-java" target="_blank" rel="noopener noreferrer" class="">Reif et al., SOAP 2018</a>).</p>
<p><strong>Framework-managed control flow.</strong> Dependency injection, annotation-driven routing, event buses and ORM callbacks mean that in a typical web application a large fraction of the entry points and edges are created by the framework at runtime. Without a model of the framework, those functions look dead. The 2022 root-cause study of JavaScript call graphs built a method to quantify which language and framework features cause missing edges, and found dynamic property accesses to be the most common root cause across its benchmarks, with the other causes varying by benchmark (<a href="https://drops.dagstuhl.de/opus/volltexte/2022/16231/" target="_blank" rel="noopener noreferrer" class="">Chakraborty et al., ECOOP 2022</a>).</p>
<p><strong>Dynamically typed languages in general.</strong> For JavaScript the state of the art is a hybrid: run a cheap dynamic pre-analysis to collect hints about object shapes and values, then feed those hints to the static phase. The technique, called approximate interpretation, produced on average 55.1% more call edges, 21.8% more reachable functions and 17.7% more resolved call sites than purely static analysis, with only 1.5% fewer monomorphic call sites — a large soundness gain at a small precision cost (<a href="https://pure.au.dk/ws/files/451742422/3656424.pdf" target="_blank" rel="noopener noreferrer" class="">Laursen, Xu and Møller, PLDI 2024</a>). Even so, an independent 2024 benchmark measured the same tool at 67.4% soundness on a JavaScript micro-benchmark, against 87.3% for a Python static tool on its own language's micro-benchmark, where soundness counts test cases passed rather than edges recovered (<a href="https://arxiv.org/pdf/2410.00603" target="_blank" rel="noopener noreferrer" class="">arXiv 2410.00603</a>). JavaScript remains the hardest ecosystem to be sound in.</p>
<p><strong>Build-dependent languages.</strong> Java and C/C++ reachability needs compiled output or a compilation database; without a successful build the analysis has a partial program. The VEX benchmark reported that a Go reachability tool requiring a build completed only 19 of its 30 cases for that reason alone (<a href="https://arxiv.org/pdf/2609.08040" target="_blank" rel="noopener noreferrer" class="">VEX-Bench, 2026</a>).</p>
<p>The engineering response to all four is the same rule: over-approximate. When a call site cannot be resolved, assume it reaches everything plausible; when an entry point set is suspect, say Unknown rather than Unreachable; when a dynamic hint is available, use it to add edges, never to remove them. The cost is more reachable verdicts to triage. The benefit is that an unreachable verdict can be waived without a second thought, which is the only thing that makes the whole exercise worth doing.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-a-trustworthy-verdict-looks-like">What a trustworthy verdict looks like<a href="https://docs.dwaar.ai/blog/reachability-analysis-why-most-dependency-cves-dont-matter#what-a-trustworthy-verdict-looks-like" class="hash-link" aria-label="Direct link to What a trustworthy verdict looks like" title="Direct link to What a trustworthy verdict looks like" translate="no">​</a></h2>
<p>A reachability verdict is only as good as what you can check about it. If you are evaluating an analysis, or building one, these are the questions whose answers decide whether an Unreachable can be waived.</p>
<ol>
<li class=""><strong>Is there a witness path for every Reachable?</strong> A verdict without the chain of functions from entry point to vulnerable symbol cannot be verified and cannot be fixed. The path is the product.</li>
<li class=""><strong>Where do the vulnerable symbols come from, and at what confidence?</strong> Authoritative ecosystem data, a human-reviewed fix-commit diff, or an automated guess. A tool that cannot answer this per advisory is guessing for you.</li>
<li class=""><strong>Is Unknown a first-class result?</strong> Count how often it appears on a real codebase. Zero is not a good sign; it means the tool has folded its uncertainty into one of the other two answers.</li>
<li class=""><strong>Which way does it err?</strong> Take a benchmark of known-reachable vulnerabilities and check for false Unreachable verdicts. The best research result is zero false negatives on a small ground truth, and other benchmarks do record misses. Treat every false Unreachable as a defect, because a single miss is a waived CVE.</li>
<li class=""><strong>Does it model your frameworks?</strong> Ask which entry points it found in a Spring, Django or Express application. If the answer is <code>main</code> and nothing else, most of the application is invisible to it.</li>
<li class=""><strong>What does it do with transitive dependencies?</strong> Analysing them per application is slow and rarely necessary; analysing them once and summarising is the approach that scales. Either way, the verdict must account for them: the 69% of flagged projects in the Maven study that inherited their vulnerability transitively are not optional.</li>
</ol>
<p>The measured reality is that package-level dependency scanning is wrong most of the time, from about three times in four to nine times in ten in the studies here, that function-level reachability corrects a large share of that while still occasionally missing a real case, and that the hardest remaining problem is not the call graph but the advisory metadata. A security program that acts on those three facts fixes the vulnerabilities that can be exploited and stops arguing about the ones that cannot.</p>]]></content:encoded>
            <category>Insights</category>
            <category>Reachability</category>
            <category>Vulnerability management</category>
            <category>Open source</category>
        </item>
        <item>
            <title><![CDATA[Do You Know Which AI Agents You Run?]]></title>
            <link>https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run</link>
            <guid>https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run</guid>
            <pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[A register of agent names will not help when a token leaks. What to record for each agent, from MCP servers to credentials, where to look, and where the case for an inventory stops.]]></description>
            <content:encoded><![CDATA[<p>Suppose a vendor tells you tomorrow that a tool connected to your AI agents was compromised and that you should treat its tokens as stolen. Which agents use that tool? Which tokens do they hold, and who can revoke them tonight? If answering takes a week of chat messages, you do not have an agent inventory, whatever your AI policy says.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-register-everyone-starts-with">The register everyone starts with<a href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run#the-register-everyone-starts-with" class="hash-link" aria-label="Direct link to The register everyone starts with" title="Direct link to The register everyone starts with" translate="no">​</a></h2>
<p>The obvious approach is a register: an intake form, a spreadsheet of approved AI tools. The <a href="https://airc.nist.gov/airmf-resources/playbook/govern/" target="_blank" rel="noopener noreferrer" class="">NIST AI RMF Playbook</a> backs the instinct. GOVERN 1.6 reads "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities", and the suggested actions include naming "a specific individual or team" to maintain the result.</p>
<p>Keep the owner. Drop the idea that a form gets you there, because a register fails in two ways.</p>
<p>First, it only knows what somebody told it. Agents get started on laptops through local MCP servers and switched on through SaaS settings, and neither path has to cross a form. I made the general case that <a class="" href="https://docs.dwaar.ai/blog/you-cannot-secure-what-you-cannot-list">a list kept by hand is always wrong</a> in April. Agents are a harder version, because starting one does not feel like deploying a system.</p>
<p>IBM calls the wider pattern shadow AI. Its <a href="https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls" target="_blank" rel="noopener noreferrer" class="">2025 Cost of a Data Breach report</a>, conducted by Ponemon Institute and based on breaches at 600 organizations, says 63% of breached organizations either had no AI governance policy or were still developing one, and that organizations with high levels of shadow AI saw an average of USD 670,000 in higher breach costs than those with a low level or none. IBM's <a href="https://www.ibm.com/think/x-force/2025-cost-of-a-data-breach-navigating-ai" target="_blank" rel="noopener noreferrer" class="">own blog post on the report</a> words the 63% more bluntly, as having no policies in place, so I am using the press release's definition. That is one study of breached organizations, sponsored and analyzed by a vendor with a commercial interest in the topic, and it measures governance. I have no study showing how many companies lack an agent list, so my claim is reasoning, not a statistic.</p>
<p>Second, a row saying "support agent, owned by the platform team" cannot answer the questions above. An agent acts through what it is connected to, and those connections change while its name stays the same.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="an-agent-is-a-bundle">An agent is a bundle<a href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run#an-agent-is-a-bundle" class="hash-link" aria-label="Direct link to An agent is a bundle" title="Direct link to An agent is a bundle" translate="no">​</a></h2>
<p>So take it apart. The model is the easy row. Record it and where it is hosted, but I consider it the least urgent, because a model name tells you little about what the agent can touch.</p>
<p>The tools tell you that. Where those tools are MCP servers, the transport matters, because it shapes where the credentials live. In the <a href="https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization" target="_blank" rel="noopener noreferrer" class="">2025-06-18 revision of the MCP specification</a>, "Authorization is OPTIONAL for MCP implementations." HTTP-based transports should follow the OAuth 2.1-based flow, while STDIO implementations are told to "instead retrieve credentials from the environment." My reading for inventory: expect a STDIO server's token in its launch environment or client configuration rather than in an OAuth grant.</p>
<p>That makes credentials the row that decides everything. The same specification warns that a stolen token "can access protected resources with requests that appear legitimate to resource servers", so know beforehand what exists and who can kill it. IBM's report points the same way: 97% of breached organizations that experienced an AI-related security incident said they lacked proper AI access controls.</p>
<p>Data follows from credentials, since what an agent can reach is mostly what its tokens allow. Ownership closes the loop. The guidance in <a href="https://csrc.nist.gov/projects/cprt/catalog#/cprt/framework/version/SP_800_53_5_2_0/home" target="_blank" rel="noopener noreferrer" class="">NIST SP 800-53 CM-8</a> lists "software owners" and "software version numbers" among the information necessary for effective accountability of system components, and enhancement CM-8(4), part of the high baseline, asks for a means of identifying the "individuals responsible and accountable for administering those components". I would hold an agent's tokens to the same standard.</p>



































<table><thead><tr><th>List</th><th>Where to find it</th><th>What to ask of it</th></tr></thead><tbody><tr><td>Agent and owner</td><td>Repositories, SaaS admin settings, developer machines</td><td>Who answers for it, who approved it, and can they turn it off?</td></tr><tr><td>Models</td><td>Agent configuration</td><td>Which model, and hosted where?</td></tr><tr><td>Tools and MCP servers</td><td>Agent configuration, repositories</td><td>STDIO or remote HTTP? Who publishes it, and which version runs?</td></tr><tr><td>Credentials</td><td>Launch environments, client configuration, secret stores, identity provider grants</td><td>What scope, when was it last rotated, who can revoke it?</td></tr><tr><td>Data reached</td><td>The scopes of the credentials above</td><td>What can it read or change with these tokens?</td></tr></tbody></table>
<p>For remote MCP servers, also ask whether the server accepts only tokens issued for itself, as the specification requires.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="two-incidents-and-what-they-do-not-show">Two incidents, and what they do not show<a href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run#two-incidents-and-what-they-do-not-show" class="hash-link" aria-label="Direct link to Two incidents, and what they do not show" title="Direct link to Two incidents, and what they do not show" translate="no">​</a></h2>
<p>Neither of these is about an agent missing from a register. They support the credential half of my argument more than the rest.</p>
<p>In August 2025, <a href="https://cloud.google.com/blog/topics/threat-intelligence/data-theft-salesforce-instances-via-salesloft-drift" target="_blank" rel="noopener noreferrer" class="">Google Threat Intelligence Group reported</a> that a threat actor it tracks as UNC6395 stole data from Salesforce instances using compromised OAuth tokens from the Salesloft Drift app. Google describes a third-party application with integrations and does not frame this as an AI agent incident. Google later found that tokens for the Drift Email integration with Google Workspace were compromised too. Its advice was to "treat any and all authentication tokens stored in or connected to the Drift platform as potentially compromised" and review all third-party integrations connected to a Drift instance. Following that requires a list of what is connected, and an agent holding tokens to your systems has the same shape.</p>
<p>In July 2025, <a href="https://aws.amazon.com/security/security-bulletins/AWS-2025-015/" target="_blank" rel="noopener noreferrer" class="">AWS disclosed</a> that an overly broad GitHub token in a build configuration let an attacker commit malicious code to the open source repository of the Amazon Q Developer extension for VS Code, affecting version 1.84.0. AWS says the code was "unsuccessful in executing due to a syntax error", and the fix was 1.85.0. Asking whether you were affected still meant knowing who had that version. So an AI coding tool's version is an inventory item, and if you build such tools, so is the token behind the build.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-stops-applying">Where this stops applying<a href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run#where-this-stops-applying" class="hash-link" aria-label="Direct link to Where this stops applying" title="Direct link to Where this stops applying" translate="no">​</a></h2>
<p>A list protects nothing. NIST says its GOVERN 1.6 actions are suggestions, and CM-8 asks for review at a frequency you define, so an inventory nobody reviews is a historical document. It gives you speed and accountability when something goes wrong. Access control is separate work.</p>
<p>A list also does not have to be complete on day one. NIST prefers inventorying all models or systems and warns that partial inventories "may not provide the value of a full inventory", but it accepts high-risk systems at minimum. If you cannot do it all at once, I would rank by what an agent holds: write access to production goes first, and a summarizer with no tokens can wait.</p>
<p>Discovery is the hard part. A self-reported register misses shadow AI, so you need identity provider, network and endpoint views alongside repositories, and each will still miss something.</p>
<p>Better standards could shrink this work. In February, <a href="https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents" target="_blank" rel="noopener noreferrer" class="">NIST's NCCoE published a concept paper</a> saying it was "interested in launching a project to demonstrate how identity standards and best practices can be applied" to software agents. I would not wait for that, because an identity standard will not tell you which agents someone started last week.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="fill-in-one-row-twice">Fill in one row, twice<a href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run#fill-in-one-row-twice" class="hash-link" aria-label="Direct link to Fill in one row, twice" title="Direct link to Fill in one row, twice" translate="no">​</a></h2>
<p>Pick the agent with the most access. Fill in its row from the table by reading its configuration, its launch environment and its token grants, without asking anyone. Then ask its owner to fill in the same row separately, and compare. Every disagreement is a mistake a register would have kept.</p>
<p>Finally, ask the owner: if every token this agent holds had to be revoked tonight, who does it, and what breaks? Write down the answer and the date. If nobody can answer, fix that before you start on agent number two.</p>
<p>Dwaar discovers what you ship from source control, registries and cloud accounts, scans it, and puts the results in one findings list with EPSS and KEV signals.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/do-you-know-which-ai-agents-you-run#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">NIST AI Resource Center, <a href="https://airc.nist.gov/airmf-resources/playbook/govern/" target="_blank" rel="noopener noreferrer" class="">AI RMF Playbook: GOVERN (GOVERN 1.6)</a>, last revised August 2023 per the <a href="https://airc.nist.gov/airmf-resources/playbook/audit-log/" target="_blank" rel="noopener noreferrer" class="">Playbook audit log</a></li>
<li class="">NIST, <a href="https://csrc.nist.gov/projects/cprt/catalog#/cprt/framework/version/SP_800_53_5_2_0/home" target="_blank" rel="noopener noreferrer" class="">SP 800-53 Release 5.2.0 catalog, CM-8</a>, issued August 27, 2025 per <a href="https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final" target="_blank" rel="noopener noreferrer" class="">NIST's publication page</a>, with the control text also in <a href="https://github.com/usnistgov/oscal-content/blob/main/nist.gov/SP800-53/rev5/json/NIST_SP-800-53_rev5_catalog.json" target="_blank" rel="noopener noreferrer" class="">NIST's OSCAL catalog</a> and CM-8(4) in its <a href="https://github.com/usnistgov/oscal-content/blob/main/nist.gov/SP800-53/rev5/json/NIST_SP-800-53_rev5_HIGH-baseline_profile.json" target="_blank" rel="noopener noreferrer" class="">high baseline profile</a></li>
<li class="">IBM, <a href="https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls" target="_blank" rel="noopener noreferrer" class="">IBM Report: 13% of Organizations Reported Breaches of AI Models or Applications, 97% of Which Reported Lacking Proper AI Access Controls</a>, July 30, 2025</li>
<li class="">IBM, <a href="https://www.ibm.com/think/x-force/2025-cost-of-a-data-breach-navigating-ai" target="_blank" rel="noopener noreferrer" class="">2025 Cost of a Data Breach Report: Navigating the AI rush without sidelining security</a>, July 30, 2025</li>
<li class="">Model Context Protocol, <a href="https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization" target="_blank" rel="noopener noreferrer" class="">Authorization, specification revision 2025-06-18</a>, June 18, 2025</li>
<li class="">Google Cloud / Google Threat Intelligence Group, <a href="https://cloud.google.com/blog/topics/threat-intelligence/data-theft-salesforce-instances-via-salesloft-drift" target="_blank" rel="noopener noreferrer" class="">Widespread Data Theft Targets Salesforce Instances via Salesloft Drift</a>, August 2025, updated August 28, 2025</li>
<li class="">Amazon Web Services, <a href="https://aws.amazon.com/security/security-bulletins/AWS-2025-015/" target="_blank" rel="noopener noreferrer" class="">Security bulletin AWS-2025-015: Amazon Q Developer for VS Code</a>, July 23, 2025</li>
<li class="">NIST, <a href="https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents" target="_blank" rel="noopener noreferrer" class="">New concept paper on identity and authority of software agents</a>, February 5, 2026</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>AI agents</category>
            <category>Inventory</category>
        </item>
        <item>
            <title><![CDATA[Stop Counting Vulnerabilities]]></title>
            <link>https://docs.dwaar.ai/blog/stop-counting-vulnerabilities</link>
            <guid>https://docs.dwaar.ai/blog/stop-counting-vulnerabilities</guid>
            <pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[A count of open findings measures backlog, not risk. Measure time to fix what is exploited or exposed, coverage of what you own and repeat findings, and know how each gets gamed.]]></description>
            <content:encoded><![CDATA[<p>Every quarter someone asks how many vulnerabilities are open, and the number goes on a slide. If it went down, the meeting moves on. Nobody in the room can say whether the findings that closed were the ones an attacker would have used, or whether the scanner behind the number looked at everything you run.</p>
<p>My position is that the open-finding count is a capacity measure being used as a risk measure, and that this steers engineers toward the wrong work. That is my argument, not a research result. The public sources I cite below are about what decides outcomes, and none of them turns on a count of open findings.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="why-the-count-feels-right-and-then-fails">Why the count feels right and then fails<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#why-the-count-feels-right-and-then-fails" class="hash-link" aria-label="Direct link to Why the count feels right and then fails" title="Direct link to Why the count feels right and then fails" translate="no">​</a></h2>
<p>The count is easy to produce, easy to chart, and it responds when people work hard. The trouble is what it rewards.</p>
<p>Take two teams as a thought experiment. One spends a sprint bumping a shared dependency, which clears a long tail of low-severity findings across many images. The other spends it moving one internet-facing service off a version with a flaw on CISA's exploited list. The first team's chart looks wonderful. The second team moved the count by one.</p>
<p>The clock the count ignores is short. The <a href="https://www.verizon.com/business/resources/reports/dbir/" target="_blank" rel="noopener noreferrer" class="">2026 Verizon DBIR</a> says 31% of breaches now start with software vulnerabilities, for incidents between 1 November 2024 and 31 October 2025. Its <a href="https://www.verizon.com/business/resources/executivebriefs/2026-dbir-executive-summary.pdf" target="_blank" rel="noopener noreferrer" class="">executive summary</a> puts credential abuse, the previous leader, at 13%. <a href="https://cloud.google.com/blog/topics/threat-intelligence/time-to-exploit-trends-2023" target="_blank" rel="noopener noreferrer" class="">Mandiant's analysis of 2023</a> put average time-to-exploit at five days, and 56% of n-days, the flaws exploited after a patch exists, were exploited within a month. A number reviewed once a quarter cannot tell you whether you beat a clock like that.</p>
<p>Incident reviews do not read like tallies either. <a href="https://www.gao.gov/products/gao-18-559" target="_blank" rel="noopener noreferrer" class="">GAO's report on the 2017 Equifax breach</a>, in which attackers accessed personal information of at least 145.5 million individuals, reports four major factors from Equifax's own investigation: identification, detection, segmenting of access to databases, and data governance. How many findings were open is not on that list.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-cisa-asks-instead">What CISA asks instead<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#what-cisa-asks-instead" class="hash-link" aria-label="Direct link to What CISA asks instead" title="Direct link to What CISA asks instead" translate="no">​</a></h2>
<p>The most useful public template I know of is a federal directive. <a href="https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">BOD 22-01</a>, from November 2021, tied deadlines to the Known Exploited Vulnerabilities catalog, with a default of two weeks for CVEs from 2021 on and six months for older ones.</p>
<p><a href="https://www.cisa.gov/news-events/directives/bod-26-04-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04</a> replaced it on 10 June 2026. Urgency now comes from four questions about a vulnerability on a specific asset. Is the asset publicly exposed? Is the CVE in KEV? Can an adversary automate every step of the exploit? Does exploitation give partial or total control? <a href="https://www.bleepingcomputer.com/news/security/cisa-tells-govt-agencies-to-patch-critical-exploited-flaws-in-3-days/" target="_blank" rel="noopener noreferrer" class="">BleepingComputer's report</a> puts the tightest deadline at as little as three days.</p>
<p>None of the four inputs is how many findings are open. The directive binds federal civilian agencies, not most readers, but I would borrow from it anyway.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="three-numbers-worth-reporting">Three numbers worth reporting<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#three-numbers-worth-reporting" class="hash-link" aria-label="Direct link to Three numbers worth reporting" title="Direct link to Three numbers worth reporting" translate="no">​</a></h2>
<p><strong>Time to fix what is exploited or exposed.</strong> Scope this to findings that are in KEV or sit on a publicly exposed asset. CISA's <a href="https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">FAQ on BOD 22-01</a> says organizations should use KEV "as an input to their vulnerability management prioritization framework", and the <a href="https://www.cisa.gov/known-exploited-vulnerabilities-catalog" target="_blank" rel="noopener noreferrer" class="">catalog page</a> carries the same sentence, so add <a href="https://www.first.org/epss/" target="_blank" rel="noopener noreferrer" class="">EPSS</a>, whose scores <a href="https://csrc.nist.gov/pubs/cswp/41/likely-exploited-vulnerabilities-a-proposed-metric/final" target="_blank" rel="noopener noreferrer" class="">NIST describes</a> as estimated probabilities that a CVE will be observed to be exploited in the wild within the next 30 days. It is a probability, not a verdict.</p>
<p>This metric lives or dies on two definitions. When does the clock start? BOD 26-04 uses the earlier of the KEV listing or the agency's own identification, and I would copy that rule. What counts as fixed? <a href="https://csrc.nist.gov/pubs/sp/800/40/r4/final" target="_blank" rel="noopener noreferrer" class="">NIST SP 800-40 Rev. 4</a> describes patch management as running through "verifying the installation of patches", a stricter bar than a closed ticket.</p>
<p>Report the oldest open item next to any average. Mandiant's own figure shows why: leaving 15 outliers in moves its average from five days to 47.</p>
<p><strong>Coverage of what you own.</strong> Every other number describes only what got scanned. So build the denominator from outside the scanner, from cloud accounts, registries and source control, and report the share that was scanned recently. BOD 26-04 points the same way by requiring agencies to attest quarterly to their publicly exposed IP addresses and domains. <a class="" href="https://docs.dwaar.ai/blog/you-cannot-secure-what-you-cannot-list">You cannot secure what you cannot list</a>, and this is where that idea turns into a number. Dwaar works on this denominator: discovering what you ship from source control, registries and cloud accounts, scanning it, and putting the results in one findings list with EPSS and KEV signals.</p>
<p><strong>Repeat findings.</strong> Count how often a closed finding comes back on the same asset. A fix that returns on the next rebuild was not a fix, and this is the one number here that says whether remediation reaches the cause. I have no public statistic on how often this happens, so treat it as my recommendation, not a benchmark.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="every-one-of-these-can-be-gamed">Every one of these can be gamed<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#every-one-of-these-can-be-gamed" class="hash-link" aria-label="Direct link to Every one of these can be gamed" title="Direct link to Every one of these can be gamed" translate="no">​</a></h2>
<p>The gaming column is my reasoning, not a study. Decide your definitions before someone with a deadline decides them for you.</p>



































<table><thead><tr><th>Metric</th><th>What it tells you</th><th>How it gets gamed</th><th>What makes that harder</th></tr></thead><tbody><tr><td>Open findings</td><td>Whether the backlog outruns your capacity</td><td>Narrow the scan scope, bulk close as accepted risk</td><td>Never report it without coverage beside it</td></tr><tr><td>Time to fix</td><td>Whether you move faster than the attacker</td><td>Start the clock at ticket creation, count a mitigation as patched, drop outliers</td><td>A written clock rule, verified fixes, the oldest open item</td></tr><tr><td>Coverage</td><td>How much of your estate the numbers describe</td><td>Shrink the denominator, count a shallow scan as covered</td><td>A denominator from systems of record, scan depth stated</td></tr><tr><td>Repeat findings</td><td>Whether fixes reach the cause</td><td>Rename, re-scope or merge findings so they no longer match</td><td>Match on the flaw and the asset, not a ticket ID</td></tr></tbody></table>
<p>Accepted risk can flatter the count and time to fix without changing anything in production. Count it separately and give it a date, because <a class="" href="https://docs.dwaar.ai/blog/a-waiver-should-expire">a waiver should expire</a>.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-stops-applying">Where this stops applying<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#where-this-stops-applying" class="hash-link" aria-label="Direct link to Where this stops applying" title="Direct link to Where this stops applying" translate="no">​</a></h2>
<p>Keep the count. A trend line in open findings tells you whether the backlog is growing, and that is a real staffing signal. Label it as capacity, not risk.</p>
<p>Time to fix has nothing to say about a flaw with no patch to time. In Mandiant's 2023 set, 97 of 138 vulnerabilities were exploited as zero-days. <a href="https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-158a" target="_blank" rel="noopener noreferrer" class="">MOVEit Transfer CVE-2023-34362</a> was a SQL injection zero-day that the CL0P ransomware gang exploited from 27 May 2023. No backlog metric could have flagged it first. What you had left was reducing exposure and responding fast.</p>
<p>KEV has limits too. An entry needs <a href="https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">a CVE ID, clear remediation guidance and reliable evidence of exploitation</a>, so it trails new attacks and misses flaws without a CVE. If KEV is your whole definition of what matters, you are measuring how fast you respond to history.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-this-on-your-own-findings">Try this on your own findings<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#try-this-on-your-own-findings" class="hash-link" aria-label="Direct link to Try this on your own findings" title="Direct link to Try this on your own findings" translate="no">​</a></h2>
<p>Export your open findings and keep only those whose CVE ID is in the KEV catalog. Next to each one, write whether its asset is publicly exposed, and its age, measured from the earlier of its KEV listing or the day you first identified it. Sort by age. The top row is a more honest headline than your total, and any row where you cannot fill in the exposure column is your coverage problem with a name on it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/stop-counting-vulnerabilities#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://www.verizon.com/business/resources/reports/dbir/" target="_blank" rel="noopener noreferrer" class="">2026 Data Breach Investigations Report</a>, Verizon Business, May 2026 (month from the page's metadata; the page shows no publication date)</li>
<li class=""><a href="https://www.verizon.com/business/resources/executivebriefs/2026-dbir-executive-summary.pdf" target="_blank" rel="noopener noreferrer" class="">2026 DBIR Executive Summary</a>, Verizon Business, June 2026</li>
<li class=""><a href="https://cloud.google.com/blog/topics/threat-intelligence/time-to-exploit-trends-2023" target="_blank" rel="noopener noreferrer" class="">Time-to-Exploit Trends: 2023</a>, Google Cloud (Mandiant), 16 October 2024</li>
<li class=""><a href="https://www.gao.gov/products/gao-18-559" target="_blank" rel="noopener noreferrer" class="">GAO-18-559: Actions Taken by Equifax and Federal Agencies</a>, U.S. GAO, 30 August 2018</li>
<li class=""><a href="https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">BOD 22-01</a>, CISA, 3 November 2021 (revoked 10 June 2026)</li>
<li class=""><a href="https://www.cisa.gov/news-events/directives/bod-26-04-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04</a>, CISA, 10 June 2026</li>
<li class=""><a href="https://www.bleepingcomputer.com/news/security/cisa-tells-govt-agencies-to-patch-critical-exploited-flaws-in-3-days/" target="_blank" rel="noopener noreferrer" class="">CISA tells govt agencies to patch critical exploited flaws in 3 days</a>, BleepingComputer, 11 June 2026</li>
<li class=""><a href="https://www.cisa.gov/known-exploited-vulnerabilities-catalog" target="_blank" rel="noopener noreferrer" class="">Known Exploited Vulnerabilities Catalog</a>, CISA, living page with no publication date</li>
<li class=""><a href="https://www.first.org/epss/" target="_blank" rel="noopener noreferrer" class="">Exploit Prediction Scoring System</a>, FIRST, living page with no publication date</li>
<li class=""><a href="https://csrc.nist.gov/pubs/cswp/41/likely-exploited-vulnerabilities-a-proposed-metric/final" target="_blank" rel="noopener noreferrer" class="">NIST CSWP 41: Likely Exploited Vulnerabilities</a>, NIST, 19 May 2025</li>
<li class=""><a href="https://csrc.nist.gov/pubs/sp/800/40/r4/final" target="_blank" rel="noopener noreferrer" class="">SP 800-40 Rev. 4</a>, NIST, April 2022</li>
<li class=""><a href="https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-158a" target="_blank" rel="noopener noreferrer" class="">AA23-158A: CL0P Exploits CVE-2023-34362 MOVEit Vulnerability</a>, CISA, 7 June 2023</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>Vulnerability management</category>
            <category>Remediation</category>
        </item>
        <item>
            <title><![CDATA[php2shell: Two Unauthenticated PHP RCEs, One Image Rebuild per Service]]></title>
            <link>https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption</link>
            <guid>https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption</guid>
            <pubDate>Thu, 22 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[CVE-2026-109618 and CVE-2026-109619 give unauthenticated RCE in PHP itself. The fix is 8.2.35, 8.3.36, 8.4.27 or 8.5.12, in every image and on every host that ships PHP.]]></description>
            <content:encoded><![CDATA[<p>You cannot fix this one from your application repository. The two flaws PHP disclosed yesterday sit in the interpreter, and the first was demonstrated over HTTP before any application code runs. The fix is a new runtime in every image and on every host, which makes today's real question an inventory question: where is all your PHP?</p>

































<table><thead><tr><th>At a glance</th><th></th></tr></thead><tbody><tr><td>Identifiers</td><td>CVE-2026-109618, <a href="https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w" target="_blank" rel="noopener noreferrer" class="">RCE via OOM corruption</a>, and CVE-2026-109619, <a href="https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp" target="_blank" rel="noopener noreferrer" class="">RCE via stale global state</a></td></tr><tr><td>Affected</td><td>PHP before 8.2.35, 8.3.36, 8.4.27 and 8.5.12, per PHP. <a href="https://www.php.net/archive/2026.php#2026-10-21-5" target="_blank" rel="noopener noreferrer" class="">PHP's announcement</a> and <a href="https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching" target="_blank" rel="noopener noreferrer" class="">CISA</a> say all PHP installations serving web requests. <a href="https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php" target="_blank" rel="noopener noreferrer" class="">UK NCSC</a> says all versions may be affected.</td></tr><tr><td>Fixed in</td><td>8.2.35, 8.3.36, 8.4.27, 8.5.12</td></tr><tr><td>Severity</td><td>Critical, CVSS 4.0 base 9.2 for both, scored by the PHP project as CNA</td></tr><tr><td>Exploited as of 2026-10-22</td><td>Attempts and early attacks reported. No confirmed compromise in these sources. Not in CISA KEV.</td></tr><tr><td>Disclosed</td><td>21 October 2026. <a href="https://www.php.net/ChangeLog-8.php" target="_blank" rel="noopener noreferrer" class="">php.net's changelog</a> dates all four fixed releases 22 October.</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-bug-is-in-how-php-gives-up">The bug is in how PHP gives up<a href="https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption#the-bug-is-in-how-php-gives-up" class="hash-link" aria-label="Direct link to The bug is in how PHP gives up" title="Direct link to The bug is in how PHP gives up" translate="no">​</a></h2>
<p>PHP handles out-of-memory errors with <code>longjmp()</code>: it jumps out of the current context and aborts the request. In some contexts, request memory is reachable before it is fully initialised. If an allocation fails there, <a href="https://nvd.nist.gov/vuln/detail/CVE-2026-109618" target="_blank" rel="noopener noreferrer" class="">request shutdown then reads or branches on invalid state</a>.</p>
<p>So the attacker's job is to exhaust memory at the right moment, and <a href="https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w" target="_blank" rel="noopener noreferrer" class="">the advisory</a> lists ways that need no application help. A multipart request without a usable declared content length can pass the initial <code>post_max_size</code> check. A part can carry unlimited headers that cost far more inside PHP than on the wire.</p>
<p>CVE-2026-109619 is the same family with a longer fuse. <a href="https://nvd.nist.gov/vuln/detail/CVE-2026-109619" target="_blank" rel="noopener noreferrer" class="">Persistent worker globals keep pointers to request memory</a> after it is freed, because a fatal error, OOM included, can abandon an operation before its matching reset. A later request on the same worker then reads or frees the stale value. PHP confirms two consumers: INI diagnostics raised before request initialisation, and session upload progress handling.</p>
<p>The CVE-2026-109618 patch removes <code>longjmp()</code> for OOM. PHP now takes a soft OOM at the next VM safe point and calls <code>_exit()</code> at a second, higher threshold, so shutdown never sees invalid state. The CVE-2026-109619 patch resets the wider request-owned state so it cannot survive into operations that run before the next request initialises.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="your-sapi-sets-the-order-not-the-exemption">Your SAPI sets the order, not the exemption<a href="https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption#your-sapi-sets-the-order-not-the-exemption" class="hash-link" aria-label="Direct link to Your SAPI sets the order, not the exemption" title="Direct link to Your SAPI sets the order, not the exemption" translate="no">​</a></h2>






























<table><thead><tr><th>SAPI</th><th>CVE-2026-109618, per <a href="https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w" target="_blank" rel="noopener noreferrer" class="">PHP's advisory</a></th><th>CVE-2026-109619, per <a href="https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp" target="_blank" rel="noopener noreferrer" class="">PHP's advisory</a></th></tr></thead><tbody><tr><td>Apache mod_php</td><td>Shutdown runs even when startup fails. RCE validated over HTTP with default input and memory limits.</td><td>RCE confirmed through the PHP 8.5 fatal backtrace path and validated through the session lifecycle path.</td></tr><tr><td>PHP-FPM</td><td>The child exits when startup fails. Source analysis suggests it stays susceptible after normal startup. No complete exploit demonstrated.</td><td>Session lifecycle RCE validated, depending on a suitable application operation. Conditionally exposed when the frontend sends <code>PHP_VALUE</code> or <code>PHP_ADMIN_VALUE</code> parameters that produce a diagnostic before request initialisation.</td></tr><tr><td>FrankenPHP</td><td>Source analysis says likely susceptible to application-independent attacks.</td><td>No equivalent early configuration consumer established in the default configuration, which PHP says does not make it immune.</td></tr><tr><td>LSAPI</td><td>Modelled on FPM. Application-specific exploits likely but unvalidated.</td><td>Can apply special environment INI values early, which makes it a conditional candidate.</td></tr></tbody></table>
<p><a href="https://bleepingcomputer.com/news/security/php-ships-emergency-patches-for-critical-php2shell-flaws" target="_blank" rel="noopener noreferrer" class="">BleepingComputer</a> quotes PHP: "We consider the risk highest for mod_php." Use the table to order your queue, not to skip anyone. Exploit portability depends on build, extensions, frontend behaviour and allocator layout, which is friction for an attacker, not a boundary for you.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-is-reported-so-far">What is reported so far<a href="https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption#what-is-reported-so-far" class="hash-link" aria-label="Direct link to What is reported so far" title="Direct link to What is reported so far" translate="no">​</a></h2>
<p><a href="https://aws.amazon.com/blogs/security/amazon-threat-intelligence-identifies-php2shell-exploitation-prior-to-public-disclosure-in-madpot/" target="_blank" rel="noopener noreferrer" class="">Amazon</a> says its MadPot honeypot logged an attempt on 17 October, four days before disclosure, and that proofs of concept appeared from about 02:00 UTC on 21 October, ahead of the 14:00 UTC embargo lift. Cloudflare CTO Dane Knecht said the company saw early attacks before the patches shipped, per BleepingComputer. <a href="https://helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619" target="_blank" rel="noopener noreferrer" class="">Help Net Security</a> reports a working exploit triggered by a single unauthenticated request.</p>
<p>Honeypot attempts are not compromises, and PHP vouches only for its own demonstrations: mod_php for <a href="https://www.openwall.com/lists/oss-security/2026/10/21/6" target="_blank" rel="noopener noreferrer" class="">CVE-2026-109618</a>, mod_php and PHP-FPM for <a href="https://www.openwall.com/lists/oss-security/2026/10/21/7" target="_blank" rel="noopener noreferrer" class="">CVE-2026-109619</a>. The missing KEV entry is still no reason to wait.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-do-now">What to do now<a href="https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption#what-to-do-now" class="hash-link" aria-label="Direct link to What to do now" title="Direct link to What to do now" translate="no">​</a></h2>
<ol>
<li class=""><strong>Upgrade.</strong> Move to 8.2.35, 8.3.36, 8.4.27 or 8.5.12, or later in the same branch, then rebuild and redeploy everything carrying the runtime. Amazon and PHP both say updating is the only complete fix.</li>
<li class=""><strong>On distribution packages, read the vendor's advisory.</strong> Help Net Security reports that Linux distributions and LTS vendors received patches before public release. Confirm the fix for the package you run.</li>
<li class=""><strong>On 8.1 or older, treat yourself as affected.</strong> PHP's advisories list only 8.2 onward and those branches get no upstream updates, but <a href="https://bleepingcomputer.com/news/security/php-ships-emergency-patches-for-critical-php2shell-flaws" target="_blank" rel="noopener noreferrer" class="">minimal source backports exist</a> for 7.4, 8.0 and 8.1, which tells you enough. NCSC says to consider taking offline what you cannot patch or mitigate.</li>
<li class=""><strong>If you cannot patch today, reduce exposure, and know it is partial.</strong> WAF rules from AWS and Cloudflare are defence in depth, not a fix.
<ul>
<li class="">For CVE-2026-109618, set and enforce <code>memory_limit=-1</code> administratively, restart every worker, and confirm that no pool, directory or application setting restores a finite limit. PHP says a cgroup <code>memory.max</code> limit can stand in for PHP's own, with caveats: it does not guarantee against allocation failure, and a kill can take the master and every worker unless each worker has its own group. Avoid <code>RLIMIT_AS</code> and <code>RLIMIT_DATA</code>, which make <code>mmap()</code> fail at the limit and reach the vulnerable path. Keep <code>post_max_size</code> conservative, and block multipart at the frontend if you do not use it.</li>
<li class="">For CVE-2026-109619, set <code>session.upload_progress.enabled=0</code> and, on 8.5, <code>fatal_error_backtraces=0</code> administratively, then restart every worker. PHP: "There is no complete generic configuration mitigation."</li>
</ul>
</li>
<li class=""><strong>Hunt from 17 October, not from disclosure.</strong> Amazon suggests looking for POST requests containing <code>PHP_SESSION_UPLOAD_PROGRESS</code> in form uploads, and for unexpected process execution or file changes on web servers. Code runs as the web worker account, so if you find signs, treat whatever that account can read as exposed.</li>
</ol>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-inventory-is-the-incident">The inventory is the incident<a href="https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption#the-inventory-is-the-incident" class="hash-link" aria-label="Direct link to The inventory is the incident" title="Direct link to The inventory is the incident" translate="no">​</a></h2>
<p>A library bug gets a dependency bump. An interpreter bug gets a new runtime under every service. Code scanning and application reachability analysis have little to say about the mod_php path, which runs before your code does, and even on FPM, where an application operation can matter, the fix is still the runtime. What decides priority is the runtime version, the SAPI and the deployment, all of which live in <a class="" href="https://docs.dwaar.ai/blog/code-scanning-is-not-enough-scan-the-artifact-you-ship">the artifact you ship</a>. Often that PHP arrived through <a class="" href="https://docs.dwaar.ai/blog/the-vulnerability-debt-you-inherit-from-your-base-image">a base image</a>, not a decision anyone remembers.</p>
<p>So do not start from a list of "our PHP services". Start from what is running: deployed images, then hosts, then software you bought. NCSC calls out PHP embedded in other software and appliances, and tells organisations to keep monitoring third-party vendors for updates to appliances and services that may bundle it.</p>
<p>Discovering what you ship from source control, registries and cloud accounts, scanning it, and keeping one findings list with EPSS and KEV signals beside it is the problem we work on at Dwaar.</p>
<p>Take your largest registry or cloud account and answer three questions for every running image: which PHP version, which SAPI, and is it reachable from the internet? Internet-facing mod_php goes first. If building that list takes longer than the patch, you have found your next project.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/php2shell-unauthenticated-php-rce-through-oom-corruption#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w" target="_blank" rel="noopener noreferrer" class="">GHSA-w8hw-w9fx-j79w, RCE via OOM corruption</a>, PHP project, 21 October 2026</li>
<li class=""><a href="https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp" target="_blank" rel="noopener noreferrer" class="">GHSA-6wmf-4rv5-p3rp, RCE via stale global state</a>, PHP project, 21 October 2026</li>
<li class=""><a href="https://www.openwall.com/lists/oss-security/2026/10/21/6" target="_blank" rel="noopener noreferrer" class="">CVE-2026-109618 announcement</a>, oss-security, 21 October 2026</li>
<li class=""><a href="https://www.openwall.com/lists/oss-security/2026/10/21/7" target="_blank" rel="noopener noreferrer" class="">CVE-2026-109619 announcement</a>, oss-security, 21 October 2026</li>
<li class=""><a href="https://www.php.net/archive/2026.php#2026-10-21-5" target="_blank" rel="noopener noreferrer" class="">Security advisory: Critical remote code execution vulnerabilities, update now</a>, PHP project, 21 October 2026</li>
<li class=""><a href="https://www.php.net/ChangeLog-8.php" target="_blank" rel="noopener noreferrer" class="">PHP 8 ChangeLog</a>, PHP project, 22 October 2026</li>
<li class=""><a href="https://nvd.nist.gov/vuln/detail/CVE-2026-109618" target="_blank" rel="noopener noreferrer" class="">CVE-2026-109618</a>, NVD, 21 October 2026</li>
<li class=""><a href="https://nvd.nist.gov/vuln/detail/CVE-2026-109619" target="_blank" rel="noopener noreferrer" class="">CVE-2026-109619</a>, NVD, 21 October 2026</li>
<li class=""><a href="https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching" target="_blank" rel="noopener noreferrer" class="">Critical Vulnerabilities in PHP Require Immediate Patching</a>, CISA, 21 October 2026</li>
<li class=""><a href="https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php" target="_blank" rel="noopener noreferrer" class="">Critical security issue affecting PHP</a>, UK NCSC, 21 October 2026</li>
<li class=""><a href="https://aws.amazon.com/blogs/security/amazon-threat-intelligence-identifies-php2shell-exploitation-prior-to-public-disclosure-in-madpot/" target="_blank" rel="noopener noreferrer" class="">Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPot</a>, AWS Security Blog, 21 October 2026</li>
<li class=""><a href="https://bleepingcomputer.com/news/security/php-ships-emergency-patches-for-critical-php2shell-flaws" target="_blank" rel="noopener noreferrer" class="">PHP ships emergency patches for critical PHP2Shell flaws</a>, BleepingComputer, 21 October 2026</li>
<li class=""><a href="https://helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619" target="_blank" rel="noopener noreferrer" class="">php2shell PHP vulnerabilities patched</a>, Help Net Security, 21 October 2026, updated 22 October 2026</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>Web frameworks</category>
            <category>Remediation</category>
        </item>
        <item>
            <title><![CDATA[CVE-2026-97229: Unauthenticated Code Execution in GitLab Through a Crafted CI/CD Regular Expression]]></title>
            <link>https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows</link>
            <guid>https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows</guid>
            <pubDate>Wed, 14 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[GitLab CVE-2026-97229 could let an unauthenticated user run code on the server via a crafted CI/CD regular expression. GitLab scores it 10.0. Fixed in 19.2.8, 19.3.5 and 19.4.2.]]></description>
            <content:encoded><![CDATA[<p>If you patched your self-managed GitLab three weeks ago for the regular expression bugs in CI/CD configuration, you still have work to do. The September releases do not cover the flaw GitLab disclosed on 12 October, and GitLab says this one can be reached without logging in.</p>

































<table><thead><tr><th>Item</th><th>Detail</th></tr></thead><tbody><tr><td>CVE</td><td><a href="https://nvd.nist.gov/vuln/detail/CVE-2026-97229" target="_blank" rel="noopener noreferrer" class="">CVE-2026-97229</a>, CWE-190 (integer overflow)</td></tr><tr><td>Affected</td><td>GitLab CE/EE 19.2 before 19.2.8, 19.3 before 19.3.5, 19.4 before 19.4.2</td></tr><tr><td>Fixed in</td><td>19.2.8, 19.3.5, 19.4.2</td></tr><tr><td>Severity</td><td>CVSS 3.1 10.0 Critical, scored by GitLab as the CNA. NVD shows no separate score of its own</td></tr><tr><td>Exploited as of 2026-10-14</td><td>No exploitation reported. Not in CISA KEV. Public PoC since 2026-10-13</td></tr><tr><td>Disclosed</td><td>2026-10-12 by GitLab. NVD published 2026-10-13</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-gitlab-has-actually-said">What GitLab has actually said<a href="https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows#what-gitlab-has-actually-said" class="hash-link" aria-label="Direct link to What GitLab has actually said" title="Direct link to What GitLab has actually said" translate="no">​</a></h2>
<p>Not much. <a href="https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">GitLab's advisory</a> names an "Integer Overflow issue in CI/CD Regular Expression Compiler". <a href="https://nvd.nist.gov/vuln/detail/CVE-2026-97229" target="_blank" rel="noopener noreferrer" class="">The description</a> says that under certain conditions it "could have allowed an unauthenticated user to execute arbitrary code on the GitLab server via a specially crafted regular expression in a CI/CD configuration, due to insufficient regular expression validation".</p>
<p>The vector is <code>CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H</code>: reachable over the network, low complexity, no privileges, no user interaction, scope changed. The reporter is joaxcar, through GitLab's HackerOne program, and it leads the <a href="https://web.archive.org/web/20261012180012/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">24 CVEs in the release</a>.</p>
<p>As a general pattern, an integer overflow in a compiler works like this: the code computes the size of what it is about to build, the arithmetic wraps in a fixed-width integer, and the buffer ends up smaller than the data written into it. That describes the bug class, not this bug. GitLab's advisory does not say how this overflow becomes code execution.</p>
<p>Two clues narrow it down.</p>
<p>The first is the <a href="https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-1-released/" target="_blank" rel="noopener noreferrer" class="">23 September release</a>. It fixed CVE-2026-89078, "a double free issue when parsing a specially crafted regular expression in a CI/CD configuration", and CVE-2026-93577, "an integer overflow issue when compiling a specially crafted regular expression in a CI/CD configuration". Both are CVSS 9.9, both need an authenticated user, and both are credited to joaxcar. The vector for CVE-2026-93577 differs from the new one in a single metric: <code>PR:L</code> instead of <code>PR:N</code>.</p>
<p>The second is the bug fix list in the <a href="https://web.archive.org/web/20261012180012/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">12 October release</a>, which carries these two lines under each of the three versions:</p>
<ul>
<li class="">"Add Ruby regexp program size patch for 3.3 and 3.4"</li>
<li class="">"Add Ruby regexp repeat-range double-free patch for 3.3 and 3.4"</li>
</ul>
<p>Both lines link to public merge requests. <a href="https://gitlab.com/gitlab-org/omnibus-gitlab/-/merge_requests/9845" target="_blank" rel="noopener noreferrer" class="">omnibus-gitlab MR 9845</a>, merged to the 19.4 stable branch on 2 October, adds the patch from <a href="https://github.com/ruby/ruby/pull/19020" target="_blank" rel="noopener noreferrer" class="">Ruby PR 19020</a>, "Avoid overflow in the regexp compiler's program size arithmetic", merged upstream on 25 September. The change is mostly in <code>regcomp.c</code>, with smaller edits in other files including <code>include/ruby/onigmo.h</code>. <a href="https://gitlab.com/gitlab-org/omnibus-gitlab/-/raw/19.4.2+ce.0/config/patches/ruby/regexp-program-size-3.4.patch" target="_blank" rel="noopener noreferrer" class="">The patch's first commit message</a> says an <code>int</code> product could overflow, so "the check that keeps the compiler from unrolling a large repeat could pass for an enormous one, hanging the compile or emitting wrapped jump offsets". <a href="https://gitlab.com/gitlab-org/omnibus-gitlab/-/merge_requests/9848" target="_blank" rel="noopener noreferrer" class="">MR 9848</a> does the same for <a href="https://github.com/ruby/ruby/pull/13332" target="_blank" rel="noopener noreferrer" class="">Ruby PR 13332</a>, a double free fix in <code>regparse.c</code> that upstream Ruby merged in May 2025.</p>
<p>GitLab itself does not tie those lines to the CVE, though public analysis places the overflow in the same regular expression compiler. The patch text describes a hang and wrapped offsets, not code execution.</p>
<p>GitLab has not said which component, value or trigger belongs to this CVE, what the "certain conditions" are, or why no authentication is needed. Outside researchers have filled in part of that. <a href="https://dailysecurityreview.com/?p=44542" target="_blank" rel="noopener noreferrer" class="">Public analysis from 13 October</a> traces the flaw to an integer overflow in Ruby's Onigmo engine, triggered when an expression that got past a US-ASCII size check is recompiled against a UTF-8 repository path through the CI lint endpoint. That is a researcher's finding, not GitLab's statement. <a href="https://www.helpnetsecurity.com/2026/10/13/gitlab-ci-cd-regular-expression-compiler-vulnerability-cve-2026-97229/" target="_blank" rel="noopener noreferrer" class="">Public testing from 13 October</a> adds a precondition to the "unauthenticated" label: the attacker must first commit a crafted .gitlab-ci.yml and a UTF-8 repository path to a branch in a public project, which an ordinary account can do on instances that allow public project creation.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-exploitation-stands">Where exploitation stands<a href="https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows#where-exploitation-stands" class="hash-link" aria-label="Direct link to Where exploitation stands" title="Direct link to Where exploitation stands" translate="no">​</a></h2>
<p>Coverage from <a href="https://www.bleepingcomputer.com/news/security/gitlab-warns-of-max-severity-rce-flaw-in-regex-compiler-engine/" target="_blank" rel="noopener noreferrer" class="">BleepingComputer</a>, <a href="https://securityaffairs.com/?p=200868" target="_blank" rel="noopener noreferrer" class="">Security Affairs</a> and <a href="https://itdaily.com/news/software/gitlab-bug-19-4-2/" target="_blank" rel="noopener noreferrer" class="">itdaily</a> carries no report of exploitation in the wild, and the CVE is not in CISA KEV. <a href="https://dailysecurityreview.com/?p=44542" target="_blank" rel="noopener noreferrer" class="">Public proof-of-concept code</a> has been available since 13 October, and <a href="https://www.helpnetsecurity.com/2026/10/13/gitlab-ci-cd-regular-expression-compiler-vulnerability-cve-2026-97229/" target="_blank" rel="noopener noreferrer" class="">researchers reproduced code execution</a> on an unpatched 19.4.1 image. No in-the-wild exploitation had been confirmed in reporting from 13 and 14 October, and that can change quickly. <a href="https://web.archive.org/web/20261014222039/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">The GitLab page as of today</a> offers no detections or indicators to help you tell whether it has reached you. A 10.0 on the system that holds your code does not need a KEV entry to justify an emergency change.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-do-now">What to do now<a href="https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows#what-to-do-now" class="hash-link" aria-label="Direct link to What to do now" title="Direct link to What to do now" translate="no">​</a></h2>
<ol>
<li class=""><strong>Find every self-managed instance and read its version.</strong> GitLab.com is already patched, and GitLab Dedicated customers need no action.</li>
<li class=""><strong>Upgrade on your minor line.</strong> 19.2 goes to 19.2.8, 19.3 goes to 19.3.5, 19.4 goes to 19.4.2. GitLab strongly recommends all self-managed installations upgrade immediately. The September versions, 19.2.7, 19.3.3 and 19.4.1, are inside the affected range.</li>
<li class=""><strong>Use 19.3.5, not 19.3.4.</strong> GitLab's <a href="https://web.archive.org/web/20261012082453/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">first version of the page</a> said "October 8" and 19.3.4. It was corrected the same day to October 12 and 19.3.5, because "19.3.4 had a bug that affected merge request pages when the Ona integration was enabled". BleepingComputer and itdaily list 19.3.4, while Security Affairs and NVD say 19.3.5. NVD's range puts 19.3.4 inside the affected versions.</li>
<li class=""><strong>Plan the outage, then take it.</strong> GitLab says "This patch will cause downtime during the upgrade as migrations must complete before GitLab can start." Multi-node installs can use the zero-downtime procedures. 19.3.5 includes a post-deploy migration.</li>
<li class=""><strong>Do not hold out for a workaround.</strong> GitLab documents none. If the upgrade has to wait a few hours, <a href="https://dailysecurityreview.com/?p=44542" target="_blank" rel="noopener noreferrer" class="">the interim guidance reported on 14 October</a> is to restrict inbound access and disable untrusted public project creation, because the attack depends on committing a crafted file to a public project branch. From that same precondition, I would also restrict anonymous access to public projects and stop untrusted users writing to them. These steps only reduce exposure along the demonstrated path and are not a fix.</li>
<li class=""><strong>Keep your logs from before the upgrade.</strong> With no published indicators, you cannot clear an instance that was exposed. If it faced the internet unpatched, decide now what you would rotate if it was touched, starting with the secrets your pipelines use.</li>
</ol>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-bug-sits-below-the-code-you-scan">The bug sits below the code you scan<a href="https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows#the-bug-sits-below-the-code-you-scan" class="hash-link" aria-label="Direct link to The bug sits below the code you scan" title="Direct link to The bug sits below the code you scan" translate="no">​</a></h2>
<p>A double free is a memory bug, and a wrapped size that ends in code execution points the same way. Ruby application code does not free memory by hand, so this sits in native code beneath the application, in a runtime GitLab packages and patches itself. You run GitLab, and GitLab runs Ruby. A pipeline configuration file is a route into native code.</p>
<p>Scanning the application's lockfile will not show you that chain, because there is no line in it for the runtime. A patch added during packaging is also easy to miss for anything that infers fixes from an upstream version number. That is the case for <a class="" href="https://docs.dwaar.ai/blog/code-scanning-is-not-enough-scan-the-artifact-you-ship">scanning the artifact you ship</a>, and the host of your pipelines deserves the same attention, because <a class="" href="https://docs.dwaar.ai/blog/your-ci-pipeline-is-production">your CI pipeline is production</a>.</p>
<p>My position is that any service compiling patterns from input it does not control is running a compiler on hostile data, and it deserves the review you would give a parser for untrusted files. Nothing I have read says whether other Ruby applications are exposed to this flaw. I would still build that list of services now.</p>
<p>Once your GitLab is upgraded, pick your service that accepts the most user-supplied configuration. Write down the runtime it really runs in production, which patches its image carries, and whether any code path compiles a regular expression from input you do not control. If you cannot answer all three, that is your next inventory gap.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/gitlab-rce-through-onigmo-regexp-integer-overflows#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">GitLab, <a href="https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">GitLab Critical Patch Release: 19.4.2, 19.3.5, 19.2.8</a>, 2026-10-12</li>
<li class="">Internet Archive, <a href="https://web.archive.org/web/20261012180012/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">GitLab page at 18:00 UTC</a>, 2026-10-12</li>
<li class="">Internet Archive, <a href="https://web.archive.org/web/20261012082453/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">first GitLab page version</a>, 2026-10-12</li>
<li class="">Internet Archive, <a href="https://web.archive.org/web/20261014222039/https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-2-released/" target="_blank" rel="noopener noreferrer" class="">GitLab page at 22:20 UTC</a>, 2026-10-14</li>
<li class="">GitLab, <a href="https://docs.gitlab.com/releases/patches/patch-release-gitlab-19-4-1-released/" target="_blank" rel="noopener noreferrer" class="">GitLab Critical Patch Release: 19.4.1, 19.3.3, 19.2.7</a>, 2026-09-23</li>
<li class="">GitLab, <a href="https://gitlab.com/gitlab-org/omnibus-gitlab/-/merge_requests/9845" target="_blank" rel="noopener noreferrer" class="">omnibus-gitlab MR 9845: Add Ruby regexp program size patch for 3.3 and 3.4 [19.4]</a>, 2026-10-02</li>
<li class="">GitLab, <a href="https://gitlab.com/gitlab-org/omnibus-gitlab/-/merge_requests/9848" target="_blank" rel="noopener noreferrer" class="">omnibus-gitlab MR 9848: Add Ruby regexp repeat-range double-free patch for 3.3 and 3.4 [19.4]</a>, 2026-10-02</li>
<li class="">GitLab, <a href="https://gitlab.com/gitlab-org/omnibus-gitlab/-/raw/19.4.2+ce.0/config/patches/ruby/regexp-program-size-3.4.patch" target="_blank" rel="noopener noreferrer" class="">regexp-program-size-3.4.patch at tag 19.4.2+ce.0</a>, 2026-10-07</li>
<li class="">Ruby, <a href="https://github.com/ruby/ruby/pull/19020" target="_blank" rel="noopener noreferrer" class="">PR 19020: Avoid overflow in the regexp compiler's program size arithmetic</a>, 2026-09-25</li>
<li class="">Ruby, <a href="https://github.com/ruby/ruby/pull/13332" target="_blank" rel="noopener noreferrer" class="">PR 13332: Prevent double free for too big repetition quantifiers</a>, 2025-05-16</li>
<li class="">NVD, <a href="https://nvd.nist.gov/vuln/detail/CVE-2026-97229" target="_blank" rel="noopener noreferrer" class="">CVE-2026-97229</a>, 2026-10-13</li>
<li class="">BleepingComputer, <a href="https://www.bleepingcomputer.com/news/security/gitlab-warns-of-max-severity-rce-flaw-in-regex-compiler-engine/" target="_blank" rel="noopener noreferrer" class="">GitLab warns of max severity code execution vulnerability</a>, 2026-10-12</li>
<li class="">Security Affairs, <a href="https://securityaffairs.com/?p=200868" target="_blank" rel="noopener noreferrer" class="">GitLab Patches Max-Severity Vulnerability CVE-2026-97229</a>, 2026-10-13</li>
<li class="">itdaily, <a href="https://itdaily.com/news/software/gitlab-bug-19-4-2/" target="_blank" rel="noopener noreferrer" class="">Critical GitLab flaw allows unauthenticated attackers to execute code</a>, 2026-10-13</li>
<li class="">Help Net Security, <a href="https://www.helpnetsecurity.com/2026/10/13/gitlab-ci-cd-regular-expression-compiler-vulnerability-cve-2026-97229/" target="_blank" rel="noopener noreferrer" class="">GitLab admins urged to patch critical RCE flaw (CVE-2026-97229)</a>, 2026-10-13</li>
<li class="">Daily Security Review, <a href="https://dailysecurityreview.com/?p=44542" target="_blank" rel="noopener noreferrer" class="">Public Exploit Released for CVSS 10.0 GitLab Flaw CVE-2026-97229</a>, 2026-10-14</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>CI/CD</category>
            <category>Open source</category>
        </item>
        <item>
            <title><![CDATA[Argo CD Repo Server Command Execution Through Kustomize]]></title>
            <link>https://docs.dwaar.ai/blog/argo-cd-repo-server-command-execution-through-kustomize</link>
            <guid>https://docs.dwaar.ai/blog/argo-cd-repo-server-command-execution-through-kustomize</guid>
            <pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Two Critical Argo CD advisories show Kustomize builds running commands in the repo-server, reachable with only Application get. Fixed in 3.3.15, 3.4.10, 3.5.4 and 3.6.0-rc2.]]></description>
            <content:encoded><![CDATA[<p>You probably treat <code>get</code> on an Argo CD Application as read access. On an unpatched server, it is enough to run commands inside <code>argocd-repo-server</code> and read the repository credentials that process can reach. The Argo CD project published two Critical advisories for this on 2026-10-07, and the fixes are already out. If you run 2.7.0 or later, upgrade this week.</p>








































<table><thead><tr><th></th><th><a href="https://github.com/argoproj/argo-cd/security/advisories/GHSA-9v9p-x54c-58gc" target="_blank" rel="noopener noreferrer" class="">Kustomize remote ref</a></th><th><a href="https://github.com/argoproj/argo-cd/security/advisories/GHSA-fw5c-w8rc-j7fx" target="_blank" rel="noopener noreferrer" class="">Kustomize Helm config home</a></th></tr></thead><tbody><tr><td>Identifier</td><td>GHSA-9v9p-x54c-58gc</td><td>GHSA-fw5c-w8rc-j7fx</td></tr><tr><td>Affected</td><td>2.7.0 up to the fixed versions, and 3.6.0-rc1</td><td>2.1.0 up to the fixed versions, and 3.6.0-rc1, only where <code>--enable-helm</code> is passed to <code>kustomize build</code></td></tr><tr><td>Fixed in</td><td><a href="https://github.com/argoproj/argo-cd/releases/tag/v3.3.15" target="_blank" rel="noopener noreferrer" class="">3.3.15</a>, <a href="https://github.com/argoproj/argo-cd/releases/tag/v3.4.10" target="_blank" rel="noopener noreferrer" class="">3.4.10</a>, <a href="https://github.com/argoproj/argo-cd/releases/tag/v3.5.4" target="_blank" rel="noopener noreferrer" class="">3.5.4</a>, <a href="https://github.com/argoproj/argo-cd/releases/tag/v3.6.0-rc2" target="_blank" rel="noopener noreferrer" class="">3.6.0-rc2</a></td><td>Same releases</td></tr><tr><td>Severity</td><td>Critical, CVSS 3.1 9.9, as published on the project's GitHub advisory</td><td>Same</td></tr><tr><td>Exploited as of 2026-10-08</td><td>No public reports in the advisories</td><td>Same</td></tr><tr><td>Disclosed</td><td>Fixed releases 2026-10-06, advisory 2026-10-07</td><td>Same</td></tr></tbody></table>
<p>Track these by GHSA id. The 9.9 is the advisory's own number, and I have no NVD score to quote. The advisories mention no public proof of concept, but they explain the mechanism well enough that I would not wait for one.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="two-bugs-both-inherited-from-kustomize">Two bugs, both inherited from Kustomize<a href="https://docs.dwaar.ai/blog/argo-cd-repo-server-command-execution-through-kustomize#two-bugs-both-inherited-from-kustomize" class="hash-link" aria-label="Direct link to Two bugs, both inherited from Kustomize" title="Direct link to Two bugs, both inherited from Kustomize" translate="no">​</a></h2>
<p>In the first, Kustomize takes a remote resource, base or component URL and <a href="https://github.com/argoproj/argo-cd/security/advisories/GHSA-9v9p-x54c-58gc" target="_blank" rel="noopener noreferrer" class="">appends its <code>ref</code> or <code>version</code> query value to <code>git fetch</code> with no <code>--</code> separator</a>. A value starting with a hyphen, such as <code>--upload-pack=&lt;command&gt;</code>, is parsed by Git as an option. That option runs the command when the remote uses a local transport such as <code>file://</code>, and not for <code>https://</code>.</p>
<p>A <code>file://</code> remote never leaves the pod, so blocking repo-server egress does not help, and Git returns the command's output in the manifest-generation error.</p>
<p>It became reachable when Argo CD's bundled Kustomize went from 4.5.7 to 5.0.1 in commit <code>25badeb7a1</code> on 2023-03-16, first released in 2.7.0. Kustomize 4.5.7 rejects <code>file://</code>, and 5.0.0 first accepts it. Upgrading Kustomize alone does not help, because current Kustomize still passes the ref without a separator. Argo CD's patch puts a Git wrapper ahead of the real <code>git</code> on the build's <code>PATH</code>, so the URL and ref cannot parse as options. Remote bases keep working.</p>
<p>The second bug needs <code>--enable-helm</code> in your build options. <a href="https://github.com/argoproj/argo-cd/security/advisories/GHSA-fw5c-w8rc-j7fx" target="_blank" rel="noopener noreferrer" class="">Kustomize copies <code>helmGlobals.configHome</code> into Helm's config and data directories</a>. Helm then loads a downloader plugin from <code>&lt;configHome&gt;/.data/plugins</code> with no separate install step, and runs its program when a <code>helmCharts</code> repository URL uses that plugin's protocol. This arrived with the Kustomize 3.9.4 to 4.1.2 upgrade in commit <code>b6d5c23bc</code> on 2021-04-20, first released in 2.1.0. The patch points Helm at an empty config and data directory for every Kustomize build.</p>
<p>Both bugs share two delivery paths:</p>
<ul>
<li class="">A Kustomization committed to a Git repository that an Application builds.</li>
<li class="">An upload to <code>POST /api/v1/applications/manifestsWithFiles</code>, the path behind <code>argocd app diff --local --server-side-generate</code>.</li>
</ul>
<p>The upload path needs only Application <code>get</code> on an existing Application, with no permission to update or sync it and no write access to its repository. Either way, the command runs as the repo-server user and can read repository credentials and anything else available to that process.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-check-and-do-now">What to check and do now<a href="https://docs.dwaar.ai/blog/argo-cd-repo-server-command-execution-through-kustomize#what-to-check-and-do-now" class="hash-link" aria-label="Direct link to What to check and do now" title="Direct link to What to check and do now" translate="no">​</a></h2>
<p><strong>Upgrade.</strong> Go to 3.3.15, 3.4.10, 3.5.4 or 3.6.0-rc2, matching your minor. Argo CD 2.x and 3.0 to 3.2 have no patch, so move those to a supported minor. If you install from the project's manifests, the <a href="https://github.com/argoproj/argo-cd/releases/tag/v3.3.15" target="_blank" rel="noopener noreferrer" class="">3.3.15 release notes</a> give this, with <code>manifests/ha/install.yaml</code> for HA:</p>
<div class="language-bash codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-bash codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token plain">kubectl apply -n argocd --server-side --force-conflicts -f https://raw.githubusercontent.com/argoproj/argo-cd/v3.3.15/manifests/install.yaml</span><br></div></code></pre></div></div>
<p><strong>Check for the breaking change first.</strong> Kustomize builds that render Helm charts no longer honor <code>helmGlobals.configHome</code>, so Helm plugins registered that way stop working.</p>
<p><strong>If you cannot upgrade today, know what each stopgap buys:</strong></p>

























<table><thead><tr><th>Stopgap</th><th>Covers</th><th>Cost or gap</th></tr></thead><tbody><tr><td>Remove <code>--enable-helm</code> from <code>kustomize.buildOptions</code></td><td>Helm bug, completely</td><td>Apps rendering Helm through Kustomize break</td></tr><tr><td>Register Kustomize 4.5.7 with <code>kustomize.path.&lt;version&gt;</code> in <code>argocd-cm</code>, and set <code>spec.source.kustomize.version</code> on every Application</td><td>Remote ref bug</td><td>Breaks apps needing Kustomize 5, and leaving 5.x as default or registered leaves it open</td></tr><tr><td>Restrict Application <code>get</code>, or block the upload endpoint</td><td>Upload path only</td><td>A Kustomization committed to Git is still built on the next sync</td></tr></tbody></table>
<p>The project says the remote ref bug has no complete workaround besides upgrading. The 4.5.7 pin also postdates Kustomize 4.1.0, which first passes <code>configHome</code> to Helm, so do not count it against the Helm bug.</p>
<p><strong>Look back.</strong> The advisories publish no indicators of compromise, so this is my reading of the mechanism, not project guidance. Search Kustomizations for remote <code>ref</code> or <code>version</code> values beginning with a hyphen, for <code>file://</code> remotes, and for <code>helmGlobals.configHome</code> alongside a <code>.data/plugins</code> directory. Review <code>manifestsWithFiles</code> calls from identities with no reason to make them, and look for command output in manifest-generation errors. On a real hit, treat the repository credentials as exposed and rotate them.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-renderer-is-part-of-your-control-plane">The renderer is part of your control plane<a href="https://docs.dwaar.ai/blog/argo-cd-repo-server-command-execution-through-kustomize#the-renderer-is-part-of-your-control-plane" class="hash-link" aria-label="Direct link to The renderer is part of your control plane" title="Direct link to The renderer is part of your control plane" translate="no">​</a></h2>
<p>A GitOps controller is trusted to turn what is in Git into what runs in a cluster, and to do that it runs third-party tools over files written by many people. Rendering is not a harmless step before the real work: it can mean running commands chosen by whoever wrote the repository.</p>
<p>The advisories stop at repository credentials and process access. My own read is that a token with write access to a repo Argo CD syncs is a short path to the cluster, so scope those tokens to read wherever you can. Stop handing out Application <code>get</code> as harmless, and treat write access to any repo an Application builds as a privileged grant, for the same reason <a class="" href="https://docs.dwaar.ai/blog/your-ci-pipeline-is-production">your CI pipeline is production</a>.</p>
<p>Both bugs came in through a bump of a bundled tool and sat there for roughly three and a half and five and a half years. An Argo CD version number does not tell you which Kustomize you run, so that belongs in your inventory too, because <a class="" href="https://docs.dwaar.ai/blog/you-cannot-secure-what-you-cannot-list">you cannot secure what you cannot list</a>. Dwaar discovers what you ship from source control, registries and cloud accounts, scans it, and keeps the results in one findings list.</p>
<p>Today, write down one line for each Argo CD instance: its version, every registered Kustomize version, whether <code>--enable-helm</code> is set, and who holds Application <code>get</code>. Then find other pipelines running <code>kustomize build</code> on repositories you do not control, because Argo CD's Git wrapper does not protect them.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/argo-cd-repo-server-command-execution-through-kustomize#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://github.com/argoproj/argo-cd/security/advisories/GHSA-9v9p-x54c-58gc" target="_blank" rel="noopener noreferrer" class="">A Kustomize remote ref can run commands in the repo-server</a>, Argo CD project, 2026-10-07</li>
<li class=""><a href="https://github.com/argoproj/argo-cd/security/advisories/GHSA-fw5c-w8rc-j7fx" target="_blank" rel="noopener noreferrer" class="">A Kustomize Helm config home can run commands in the repo-server</a>, Argo CD project, 2026-10-07</li>
<li class=""><a href="https://github.com/argoproj/argo-cd/releases/tag/v3.3.15" target="_blank" rel="noopener noreferrer" class="">Argo CD v3.3.15 release</a>, Argo CD project, 2026-10-06</li>
<li class=""><a href="https://github.com/argoproj/argo-cd/releases/tag/v3.4.10" target="_blank" rel="noopener noreferrer" class="">Argo CD v3.4.10 release</a>, Argo CD project, 2026-10-06</li>
<li class=""><a href="https://github.com/argoproj/argo-cd/releases/tag/v3.5.4" target="_blank" rel="noopener noreferrer" class="">Argo CD v3.5.4 release</a>, Argo CD project, 2026-10-06</li>
<li class=""><a href="https://github.com/argoproj/argo-cd/releases/tag/v3.6.0-rc2" target="_blank" rel="noopener noreferrer" class="">Argo CD v3.6.0-rc2 release</a>, Argo CD project, 2026-10-06</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>Containers</category>
            <category>CI/CD</category>
        </item>
        <item>
            <title><![CDATA[A Waiver Should Expire]]></title>
            <link>https://docs.dwaar.ai/blog/a-waiver-should-expire</link>
            <guid>https://docs.dwaar.ai/blog/a-waiver-should-expire</guid>
            <pubDate>Tue, 06 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Risk acceptance with no end date is a hidden backlog. What a waiver should record, what should reopen it, and why suppressing noise is a different decision.]]></description>
            <content:encoded><![CDATA[<p>Filter your findings list to everything marked accepted and sort by age. Who approved the oldest one, what were they told, and is any of it still true? If the record cannot answer those questions, you do not have accepted risk. You have a backlog nobody is required to look at.</p>
<p>The route here is honest. A finding cannot be fixed this sprint, someone has a good reason, and the tracker offers a status called accepted with a comment box. The decision may have been right that day. The record keeps asserting it long after.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="why-the-status-field-fails">Why the status field fails<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#why-the-status-field-fails" class="hash-link" aria-label="Direct link to Why the status field fails" title="Direct link to Why the status field fails" translate="no">​</a></h2>
<p>A status field records an outcome and removes the item from the view people work from. Nothing brings it back.</p>
<p>NIST's Risk Management Framework (RMF) shows what that status leaves out. SP 800-37 says plan of action entries <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-37r2.pdf" target="_blank" rel="noopener noreferrer" class="">"are not necessary when deficiencies are accepted by the authorizing official as residual risk."</a> Fair enough, since a plan of action <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-53r5.pdf" target="_blank" rel="noopener noreferrer" class="">documents planned remediation</a> and acceptance means there is none. NIST does not let the item vanish, though: accepted deficiencies <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-37r2.pdf" target="_blank" rel="noopener noreferrer" class="">"remain documented in the security and privacy assessment reports and are monitored for changes to the risk factors."</a> A status with a comment box copies the exemption, so the item leaves the working list, and it skips the monitoring.</p>
<p>The RMF also gives the authorization itself a termination date that "indicates when the authorization expires." NIST does not apply that date to individual accepted findings, and the RMF is written for authorizing systems, not for waiving one CVE on one image. So I am borrowing its shape, not quoting a rule that binds your tracker.</p>
<p>The shape is right because the facts under a waiver move. CISA's <a href="https://www.cisa.gov/news-events/directives/bod-26-04-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04</a>, issued on 10 June 2026 to replace BOD 22-01 and BOD 19-02, says so directly: "The timelines for each asset are dynamic as facts change." Its inputs are KEV status, public exposure, automatability and technical impact, and a new KEV listing shortens the timeline.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="two-decisions-hiding-under-one-button">Two decisions hiding under one button<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#two-decisions-hiding-under-one-button" class="hash-link" aria-label="Direct link to Two decisions hiding under one button" title="Direct link to Two decisions hiding under one button" translate="no">​</a></h2>
<p>Before expiry can mean anything, separate noise from acceptance.</p>
<p>Suppression is a claim of fact: this does not apply here. A waiver is a decision: this applies, and we are living with it for now. CISA's <a href="https://www.cisa.gov/sites/default/files/2023-01/VEX_Status_Justification_Jun22.pdf" target="_blank" rel="noopener noreferrer" class="">VEX status justifications</a> show the first kind done properly. A not affected status can carry a machine-readable reason in an optional field, and the document lists five, such as <code>Component_not_present</code> or <code>Vulnerable_code_not_in_execute_path</code>. Even then, CISA says the consumer "may choose to accept it or not" and that these statuses are "not intended to be a discussion-ending declaration." A suppression is <a class="" href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof">a claim like any finding</a>, and it can be wrong.</p>

























<table><thead><tr><th></th><th>Suppression</th><th>Waiver</th></tr></thead><tbody><tr><td>Who makes the call</td><td>An engineer who can show the evidence</td><td>Someone accountable for the risk</td></tr><tr><td>What backs it</td><td>A fact, such as the component being absent</td><td>A rationale and any control relied on</td></tr><tr><td>What ends it</td><td>The code or the reachability changes</td><td>A date passes or a named condition changes</td></tr></tbody></table>
<p>A single status for both does damage each way. Waivers get buried in test fixture noise, and real exposures can be filed as false positives because that path asks for no approver.</p>
<p>The border case is CISA's fifth justification, <code>Inline_mitigations_already_exist</code>. My position: if the reason depends on a control someone has to keep running, it is a waiver. Give it an owner and a date.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-the-record-has-to-hold">What the record has to hold<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#what-the-record-has-to-hold" class="hash-link" aria-label="Direct link to What the record has to hold" title="Direct link to What the record has to hold" translate="no">​</a></h2>
<p>Who, why, until when, and under what condition. An illustration, not a schema from any standard or product:</p>
<div class="language-yaml codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-yaml codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token comment"># Illustration only.</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">finding</span><span class="token punctuation">:</span><span class="token plain"> CVE</span><span class="token punctuation">-</span><span class="token plain">YYYY</span><span class="token punctuation">-</span><span class="token plain">NNNNN</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">scope</span><span class="token punctuation">:</span><span class="token plain"> one image in one environment</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">accepted_by</span><span class="token punctuation">:</span><span class="token plain"> a named person with authority over this risk</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">rationale</span><span class="token punctuation">:</span><span class="token plain"> why it is not being fixed now</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">relies_on</span><span class="token punctuation">:</span><span class="token plain"> the control in place</span><span class="token punctuation">,</span><span class="token plain"> and how we know it is running</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">expires</span><span class="token punctuation">:</span><span class="token plain"> YYYY</span><span class="token punctuation">-</span><span class="token plain">MM</span><span class="token punctuation">-</span><span class="token plain">DD</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token key atrule">reopen_when</span><span class="token punctuation">:</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token punctuation">-</span><span class="token plain"> the CVE is added to KEV</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token punctuation">-</span><span class="token plain"> the asset becomes publicly exposed</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token punctuation">-</span><span class="token plain"> a public exploit appears</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token punctuation">-</span><span class="token plain"> a fix ships for the version we run</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token punctuation">-</span><span class="token plain"> the control above changes or stops</span><br></div></code></pre></div></div>
<p>For the who, the RMF is blunt: <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-37r2.pdf" target="_blank" rel="noopener noreferrer" class="">"the authorizing official is the only person who can accept risk."</a> Keep the principle: one named person with authority over the risk, not a team alias, and not the engineer who wants the ticket closed.</p>
<p>For the why, SP 800-53 RA-7 lists <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-53r5.pdf" target="_blank" rel="noopener noreferrer" class="">"accepting risk with appropriate justification or rationale"</a> as one response among several. "Low priority" is not a justification. Keep the scope narrow too: BOD 26-04 sets timelines per asset, and a waiver for a CVE everywhere is a policy nobody reviewed.</p>
<p>A date catches what you failed to predict. Conditions catch what you did predict. A waiver with only a date sits unchanged through a KEV listing, and one with only conditions lasts forever if the conditions are wrong. BOD 26-04's inputs are a good start for the list, and <a class="" href="https://docs.dwaar.ai/blog/cvss-is-not-risk-how-to-use-epss-and-kev">CVSS Is Not Risk</a> covers why KEV belongs on it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="expiry-has-to-be-loud">Expiry has to be loud<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#expiry-has-to-be-loud" class="hash-link" aria-label="Direct link to Expiry has to be loud" title="Direct link to Expiry has to be loud" translate="no">​</a></h2>
<p>A date alone is not a control. When it passes, I want the finding back in the open list at its real severity, not a reminder email someone can archive.</p>
<p>The <a href="https://oversight.house.gov/wp-content/uploads/2018/12/Equifax-Report.pdf" target="_blank" rel="noopener noreferrer" class="">House Oversight majority staff report on Equifax</a> shows what a silent lapse looks like. The report does not describe a waiver, but "the device used to monitor ACIS network traffic had been inactive for 19 months due to an expired security certificate", and it stayed that way until Equifax updated the certificate on 29 July 2017. An internal tracker entry dated 20 January 2017 had already recorded the wider problem: "SSLV devices are missing certificates, limiting visibility to web based attacks". Something expired, the wider problem was written down, and nothing forced a response. That is why <code>relies_on</code> is in the illustration: a waiver resting on a control is only as good as the last check of that control.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-position-stops-applying">Where this position stops applying<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#where-this-position-stops-applying" class="hash-link" aria-label="Direct link to Where this position stops applying" title="Direct link to Where this position stops applying" translate="no">​</a></h2>
<p>Ongoing authorization is the first limit. NIST says organizations "may eliminate the authorization termination date" when a system operates under ongoing authorization, but that route swaps the date for a review frequency and continued, communicated acceptance by the official. Expiry becomes review, not permanence. If your conditions are evaluated continuously and someone reaffirms on a schedule, drop the date.</p>
<p>Deadlines you do not own are the second. A waiver in this post means internal acceptance. CISA's <a href="https://www.cisa.gov/news-events/directives/bod-26-04-implementation-guidance-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">implementation guidance</a> states that "CISA does not issue waivers or exceptions for actions required in cyber directives." If someone else owns the deadline, your waiver records your decision and moves nothing.</p>
<p>The third is the rubber stamp. Expiring waivers create work, and I have no study on how renewals behave, so this is my argument: renewal means answering the conditions again, not typing a new date. If nobody can answer, the finding goes back to open.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-this-on-your-own-list">Try this on your own list<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#try-this-on-your-own-list" class="hash-link" aria-label="Direct link to Try this on your own list" title="Direct link to Try this on your own list" translate="no">​</a></h2>
<p>Export every finding marked accepted, ignored or false positive, and take the oldest ten. Sort them into the two columns of the table above. For each waiver, check two things today: is the CVE in the KEV catalog now, and is the asset publicly exposed now? CISA's test works as written: accessible to unauthenticated or untrusted entities via public networks. Unless that was already true when the waiver was signed, a yes means the decision rests on old facts, so reopen it. Everything else gets one named owner and one date this week.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/a-waiver-should-expire#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">NIST, <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-37r2.pdf" target="_blank" rel="noopener noreferrer" class="">SP 800-37 Rev. 2, Risk Management Framework</a>, December 2018</li>
<li class="">NIST, <a href="https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-53r5.pdf" target="_blank" rel="noopener noreferrer" class="">SP 800-53 Rev. 5, RA-7 and CA-5</a>, September 2020, includes updates as of 10 December 2020</li>
<li class="">CISA, <a href="https://www.cisa.gov/sites/default/files/2023-01/VEX_Status_Justification_Jun22.pdf" target="_blank" rel="noopener noreferrer" class="">Vulnerability Exploitability eXchange (VEX): Status Justifications</a>, June 2022</li>
<li class="">CISA, <a href="https://www.cisa.gov/news-events/directives/bod-26-04-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04: Prioritizing Security Updates Based on Risk</a>, 10 June 2026</li>
<li class="">CISA, <a href="https://www.cisa.gov/news-events/directives/bod-26-04-implementation-guidance-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04 Implementation Guidance</a>, updated 25 August 2026</li>
<li class="">US House Oversight Committee majority staff, <a href="https://oversight.house.gov/wp-content/uploads/2018/12/Equifax-Report.pdf" target="_blank" rel="noopener noreferrer" class="">The Equifax Data Breach</a>, December 2018</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>Vulnerability management</category>
            <category>Remediation</category>
        </item>
        <item>
            <title><![CDATA[CVE-2026-94603: A Podman Image Annotation Switches Off the Sandbox You Asked For]]></title>
            <link>https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox</link>
            <guid>https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox</guid>
            <pubDate>Thu, 01 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[On rootful Podman from 4.4.0, one image annotation makes podman run ignore flags such as --cap-drop=ALL. Podman 5.8.8 and 6.1.3 remove the feature.]]></description>
            <content:encoded><![CDATA[<p>You do not fully trust an image, so you run it with <code>--cap-drop=ALL</code> and move on. On rootful Podman from 4.4.0 until this week's fixes, that image could overrule you with one annotation, and Podman would <a href="https://github.com/podman-container-tools/podman/security/advisories/GHSA-2cvf-wqm6-wr9g" target="_blank" rel="noopener noreferrer" class="">silently ignore</a> what you asked for.</p>

































<table><thead><tr><th>Item</th><th>Detail</th></tr></thead><tbody><tr><td>Identifier</td><td>CVE-2026-94603, GHSA-2cvf-wqm6-wr9g</td></tr><tr><td>Affected</td><td>Podman from 4.4.0 before the fixed releases, rootful containers only</td></tr><tr><td>Fixed in</td><td>5.8.8 and 6.1.3</td></tr><tr><td>Severity</td><td>10.0 Critical, CVSS 3.1, as shown on the Podman project's GitHub advisory</td></tr><tr><td>Exploited as of 2026-10-01</td><td>None reported in the advisory or the release notes</td></tr><tr><td>Disclosed</td><td>2026-09-29</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="an-image-and-a-checkpoint-are-not-the-same-thing">An image and a checkpoint are not the same thing<a href="https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox#an-image-and-a-checkpoint-are-not-the-same-thing" class="hash-link" aria-label="Direct link to An image and a checkpoint are not the same thing" title="Direct link to An image and a checkpoint are not the same thing" translate="no">​</a></h2>
<p>Podman 4.4.0 <a href="https://github.com/podman-container-tools/podman/security/advisories/GHSA-2cvf-wqm6-wr9g" target="_blank" rel="noopener noreferrer" class="">added a feature</a> that let checkpoints from <code>podman container checkpoint</code> travel through an OCI registry. The <a href="https://github.com/podman-container-tools/podman/commit/f4401567cdf6d7ccc1ef9f50345c865bc6262c22" target="_blank" rel="noopener noreferrer" class="">original commit</a>, written in November 2022, states the behaviour plainly: "When <code>podman run</code> is invoked with an image that contains a checkpoint, it would restore the container from that checkpoint."</p>
<p>That made one command mean two things. Running an image means you decide the sandbox. Restoring a checkpoint means the checkpoint brings its own container configuration. The thing that chose between those two modes was the image itself, through one line in the code that the <a href="https://github.com/podman-container-tools/podman/commit/3721ff82ffb44a4ceeaecc5abd3808e924a6b669" target="_blank" rel="noopener noreferrer" class="">fix removed</a>:</p>
<div class="language-go codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-go codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token boolean">_</span><span class="token punctuation">,</span><span class="token plain"> isCheckpointImage </span><span class="token operator">:=</span><span class="token plain"> imgData</span><span class="token punctuation">.</span><span class="token plain">Annotations</span><span class="token punctuation">[</span><span class="token plain">define</span><span class="token punctuation">.</span><span class="token plain">CheckpointAnnotationRuntimeName</span><span class="token punctuation">]</span><br></div></code></pre></div></div>
<p>That constant is the key <code>io.podman.annotations.checkpoint.runtime.name</code>. The advisory is blunt about the test: "If this annotation is set, the image is treated as a checkpoint." Presence of the key was enough.</p>
<p>When the key was present, <code>podman run</code> called <code>ic.ContainerRestore</code> with options carrying only the name and pod from the run spec, then returned a report straight away. That return sits before <code>generate.MakeContainer</code> in the same function, so your flags never became a container spec. The early return is my reading of the diff. The advisory states the outcome directly: Podman will "ignore --cap-drop=all and use the set of capabilities specified by the checkpoint image instead, which could include all capabilities."</p>
<p>The project's sources differ on how far this goes. The advisory says "almost all sandboxing". The <a href="https://github.com/podman-container-tools/podman/releases/tag/v5.8.8" target="_blank" rel="noopener noreferrer" class="">5.8.8 release notes</a> say it "could disable all sandboxing, including sandboxing specified by the user". Plan for the second wording.</p>
<p>The maintainers name the root issue in one clause: this "allows inherently insecure behavior as the image can now control how it is executed." This is a design flaw, not a memory bug, so there was no parser to harden. They <a href="https://github.com/podman-container-tools/podman/security/advisories/GHSA-2cvf-wqm6-wr9g" target="_blank" rel="noopener noreferrer" class="">removed the feature</a>, <a href="https://github.com/podman-container-tools/podman/commit/3721ff82ffb44a4ceeaecc5abd3808e924a6b669" target="_blank" rel="noopener noreferrer" class="">deleting 64 lines</a> across the local and remote code paths in the revert, then <a href="https://github.com/podman-container-tools/podman/commit/c74b1d37b3fb834b0de43e176565d9d8c254b627" target="_blank" rel="noopener noreferrer" class="">added a server-side check</a>, so <code>podman run</code> on an annotated image now fails with an error telling you to use <code>podman container restore</code>. I think that was the right call. Any stricter check on the annotation would still be the image deciding how it runs.</p>
<p>The advisory credits ADA Logics Ltd, using Anthropic's Claude models, with the discovery.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="read-the-100-against-your-own-setup">Read the 10.0 against your own setup<a href="https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox#read-the-100-against-your-own-setup" class="hash-link" aria-label="Direct link to Read the 10.0 against your own setup" title="Direct link to Read the 10.0 against your own setup" translate="no">​</a></h2>
<p>The advisory's vector is <code>CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H</code>. It claims no user interaction, yet someone or something has to run the image. That sits awkwardly, and it is why <a class="" href="https://docs.dwaar.ai/blog/cvss-is-not-risk-how-to-use-epss-and-kev">a score alone does not set your priority</a>.</p>
<p>My position is to rank by one question: can an image someone else built reach a rootful <code>podman run</code> on your systems? CI runners and automation that pull third-party images go first. A host running only images you build yourself can follow in normal patching. Neither the advisory nor the release notes report exploitation, but the disclosure is only two days old.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-check-and-do-now">What to check and do now<a href="https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox#what-to-check-and-do-now" class="hash-link" aria-label="Direct link to What to check and do now" title="Direct link to What to check and do now" translate="no">​</a></h2>
<ol>
<li class=""><strong>Find the exposed hosts.</strong> You are looking for rootful Podman at 4.4.0 or later. Rootless containers <a href="https://github.com/podman-container-tools/podman/security/advisories/GHSA-2cvf-wqm6-wr9g" target="_blank" rel="noopener noreferrer" class="">are not affected</a>, because they "do not support checkpoint/restore at all currently".</li>
<li class=""><strong>Upgrade to 5.8.8 on the 5.x line or 6.1.3 on the 6.x line, or later.</strong> The advisory lists <code>github.com/containers/podman</code> patched at v5.8.8 and <code>go.podman.io/podman/v6</code> patched at v6.1.3, so check Go projects that import those modules as well. Upstream names no other fixed release. If you run a distribution package, ask your distribution whether it carries the revert instead of judging by the version number. The remote client held its own copy of the check, so upgrade <code>podman-remote</code> clients along with hosts.</li>
<li class=""><strong>Until then, apply the advisory's workaround.</strong> Scan pulled images for <code>io.podman.annotations.checkpoint.runtime.name</code> and reject any image that has it. The advisory says Podman "does not presently have the ability to perform this scanning and rejection automatically", so this check has to live in your own pipeline, at a registry mirror or a CI gate. The advisory gives no exact command. The vulnerable code read the annotation from the image's inspect data, so <code>podman image inspect</code> output is where to look on a host. Do not run untrusted images under rootful <code>podman run</code> until this is in place.</li>
<li class=""><strong>Look backwards.</strong> Search registry mirrors, CI caches and hosts for images carrying the key. Then compare the capabilities of running containers against what you requested. That second check is my inference from the described behaviour, not a detection any source documents.</li>
<li class=""><strong>If you used this feature, expect breakage.</strong> The <a href="https://github.com/podman-container-tools/podman/releases/tag/v5.8.8" target="_blank" rel="noopener noreferrer" class="">release notes</a> list it as a breaking change: "Removed support for checkpoint images in podman run due to serious security concerns". On the fixed releases, <code>podman run</code> on such an image <a href="https://github.com/podman-container-tools/podman/commit/c74b1d37b3fb834b0de43e176565d9d8c254b627" target="_blank" rel="noopener noreferrer" class="">now errors</a> rather than starting a container, and the message says the image "must be started using <code>podman container restore</code> instead". Move to that explicit command. The <a href="https://github.com/podman-container-tools/podman/commit/3721ff82ffb44a4ceeaecc5abd3808e924a6b669" target="_blank" rel="noopener noreferrer" class="">revert commit</a> says checkpoints "must be explicitly requested by the user" and recommends monitoring a checkpoint's whole lifecycle and ensuring its integrity at each stage.</li>
</ol>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="image-metadata-is-attacker-input">Image metadata is attacker input<a href="https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox#image-metadata-is-attacker-input" class="hash-link" aria-label="Direct link to Image metadata is attacker input" title="Direct link to Image metadata is attacker input" translate="no">​</a></h2>
<p>A review of an image's packages has nothing to report here, because the package list can be spotless while the annotation sits beside it. A source scan is looking at a different object altogether. You have to <a class="" href="https://docs.dwaar.ai/blog/code-scanning-is-not-enough-scan-the-artifact-you-ship">scan the artifact you ship</a>, and that includes its metadata, not only its contents. The advisory's only workaround is an artifact check, which tells you where the control has to sit.</p>
<p>Treat every field in an image as written by whoever built it, and for each tool that reads a field, ask whether that field changes what the tool does. Instructions about how to run a container should come from the operator. Image data should never be able to widen them. If you build tooling that reads annotations, do not let one loosen a security setting. Make the dangerous mode something the user explicitly asks for, which is where Podman has now landed.</p>
<p>Pick one host running rootful Podman today. List the images it has pulled, read the annotations on each, and write down which keys you cannot explain. Start your reject rule with the key above, and grow it from that list.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/podman-cve-2026-94603-an-image-annotation-disables-the-sandbox#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">Podman project, <a href="https://github.com/podman-container-tools/podman/security/advisories/GHSA-2cvf-wqm6-wr9g" target="_blank" rel="noopener noreferrer" class="">GHSA-2cvf-wqm6-wr9g: The <code>podman run</code> command can be instructed to disable almost all sandboxing by image annotation</a>, 2026-09-29</li>
<li class="">Podman project, <a href="https://github.com/podman-container-tools/podman/releases/tag/v5.8.8" target="_blank" rel="noopener noreferrer" class="">Release v5.8.8</a>, 2026-09-29</li>
<li class="">Podman project, <a href="https://github.com/podman-container-tools/podman/releases/tag/v6.1.3" target="_blank" rel="noopener noreferrer" class="">Release v6.1.3</a>, 2026-09-29</li>
<li class="">Podman project, <a href="https://github.com/podman-container-tools/podman/commit/3721ff82ffb44a4ceeaecc5abd3808e924a6b669" target="_blank" rel="noopener noreferrer" class="">Revert of "Enable podman run for checkpoint images"</a>, 2026-09-28</li>
<li class="">Podman project, <a href="https://github.com/podman-container-tools/podman/commit/c74b1d37b3fb834b0de43e176565d9d8c254b627" target="_blank" rel="noopener noreferrer" class="">Error when a checkpoint image is run</a>, 2026-09-28</li>
<li class="">Podman project, <a href="https://github.com/podman-container-tools/podman/commit/f4401567cdf6d7ccc1ef9f50345c865bc6262c22" target="_blank" rel="noopener noreferrer" class="">Enable podman run for checkpoint images (original feature commit)</a>, 2022-11-20</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>Containers</category>
            <category>Artifact scanning</category>
        </item>
        <item>
            <title><![CDATA[MCP SDK OAuth Clients Could Send Credentials to a Server-Chosen Authorization Server]]></title>
            <link>https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer</link>
            <guid>https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer</guid>
            <pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[The official MCP Python and TypeScript OAuth clients let a connected server steer credentials to an authorization server it chose. Upgrading alone does not finish the fix.]]></description>
            <content:encoded><![CDATA[<p>Your agent holds a client secret for a real authorization server, and it connects to an MCP server. Which of the two decides where that secret goes? On affected versions of the official MCP SDKs, the MCP server could. The <a href="https://github.com/modelcontextprotocol/python-sdk/security/advisories/GHSA-qx49-fqc8-xw99" target="_blank" rel="noopener noreferrer" class="">Python</a> and <a href="https://github.com/modelcontextprotocol/typescript-sdk/security/advisories/GHSA-6qxp-vccf-f47h" target="_blank" rel="noopener noreferrer" class="">TypeScript</a> advisories published this week say so in their titles, and bumping the version does not finish the job.</p>








































<table><thead><tr><th>At a glance</th><th>Python SDK</th><th>TypeScript SDK</th></tr></thead><tbody><tr><td>Identifier</td><td>GHSA-qx49-fqc8-xw99</td><td>GHSA-6qxp-vccf-f47h</td></tr><tr><td>Affected</td><td><code>mcp</code> <code>&gt;= 1.9.1, &lt; 1.30.0</code> and <code>&gt;= 2.0.0a1, &lt; 2.2.0</code></td><td><code>@modelcontextprotocol/sdk</code> <code>&gt;= 1.12.0, &lt; 1.31.0</code>; <code>@modelcontextprotocol/client</code> <code>&gt;= 2.0.0, &lt; 2.2.0</code></td></tr><tr><td>Fixed in</td><td>1.30.0, 2.2.0</td><td>sdk 1.31.0, client 2.2.0</td></tr><tr><td>Severity</td><td>7.5 High (CVSS 3.1) in the project's GitHub advisory, no separate scorer named</td><td>7.5 High (CVSS 3.1) in the project's GitHub advisory, no separate scorer named</td></tr><tr><td>Exploited as of 2026-09-30</td><td>No public report found; not in CISA KEV</td><td>No public report found; not in CISA KEV</td></tr><tr><td>Disclosed</td><td>2026-09-28; fix released 2026-09-07</td><td>2026-09-30; fix released 2026-09-28</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="two-defects-and-not-every-version-has-both">Two defects, and not every version has both<a href="https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer#two-defects-and-not-every-version-has-both" class="hash-link" aria-label="Direct link to Two defects, and not every version has both" title="Direct link to Two defects, and not every version has both" translate="no">​</a></h2>
<p><strong>The metadata issuer was not always checked.</strong> The Python client accepted authorization server metadata without comparing its <code>issuer</code> to the URL the well-known document was built from. The <a href="https://github.com/modelcontextprotocol/python-sdk/security/advisories/GHSA-qx49-fqc8-xw99" target="_blank" rel="noopener noreferrer" class="">Python advisory</a> says 1.9.1 through 1.29.1 had "no <code>issuer</code> check and no credential binding on any path". Versions 2.0.0 through 2.1.1 lacked both only on the legacy fallback, used when a server publishes no protected resource metadata (PRM), and on a 403 <code>insufficient_scope</code> challenge. The advisory's range starts at 2.0.0a1 while its prose starts at 2.0.0, so treat the pre-releases as affected. The TypeScript advisory does not list this check as a defect, and the two TypeScript lines differ: <a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2888" target="_blank" rel="noopener noreferrer" class="">1.x does not compare the metadata issuer with the URL it was fetched from</a>, while <a href="https://github.com/modelcontextprotocol/typescript-sdk/blob/v2.2.0/packages/client/src/client/auth.ts" target="_blank" rel="noopener noreferrer" class="">2.x throws <code>IssuerMismatchError</code></a> unless <code>skipIssuerMetadataValidation</code> is set.</p>
<p><strong>Credentials were not bound to an issuer.</strong> Both advisories describe this defect. On the 1.x lines, stored or pre-provisioned client IDs, client secrets and refresh tokens carried no record of who issued them, so they went to whichever authorization server the resource advertised next, or on the 2025-03-26 fallback <a href="https://github.com/modelcontextprotocol/python-sdk/pull/3431" target="_blank" rel="noopener noreferrer" class="">to the resource server's own origin</a>. TypeScript 2.x already had a binding, and its advisory limits 2.x exposure to gaps such as bundled providers without <code>expectedIssuer</code> and credentials stored without <code>issuer</code>. Machine-to-machine providers hold credentials the SDK never obtained, and as <a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2887" target="_blank" rel="noopener noreferrer" class="">TypeScript PR #2887</a> puts it, "nothing but <code>expectedIssuer</code> says which authorization server they belong to". The Python equivalents had the same gap: <code>ClientCredentialsOAuthProvider</code> and <code>PrivateKeyJWTOAuthProvider</code> "used whichever authorization server discovery produced; nothing else in the flow can say which server issued those credentials", in the words of <a href="https://github.com/modelcontextprotocol/python-sdk/pull/3398" target="_blank" rel="noopener noreferrer" class="">Python PR #3398</a>.</p>
<p>The Python advisory describes two routes. A malicious or compromised server can name its own authorization server in its PRM. Or it can publish no PRM and serve metadata presenting your real authorization server as issuer, and its token endpoint "would then receive the client_secret, authorization code and PKCE code_verifier meant for that real server". With <code>PrivateKeyJWTOAuthProvider</code>, it gets a signed client assertion.</p>
<p>The <a href="https://github.com/modelcontextprotocol/typescript-sdk/security/advisories/GHSA-6qxp-vccf-f47h" target="_blank" rel="noopener noreferrer" class="">TypeScript advisory</a> adds the unattended case: a <code>refresh_token</code> and <code>client_secret</code> from an earlier sign-in, or a bundled provider's secret or assertion, could leave with no user interaction.</p>
<p>The fixes differ too. The Python fix decides the expected issuer before fetching metadata, rejects a mismatch and binds stored registrations to that issuer. The TypeScript fix records the authorization server as <code>issuer</code> on saved credentials and does not send them to a different one, but <a href="https://github.com/modelcontextprotocol/typescript-sdk/blob/1.31.0/src/client/auth.ts" target="_blank" rel="noopener noreferrer" class=""><code>@modelcontextprotocol/sdk</code> 1.31.0</a> still does not compare the metadata <code>issuer</code> with the URL it was fetched from.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="are-you-exposed">Are you exposed?<a href="https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer#are-you-exposed" class="hash-link" aria-label="Direct link to Are you exposed?" title="Direct link to Are you exposed?" translate="no">​</a></h2>
<p>You need all of this: an HTTP MCP client, the SDK's OAuth client, credentials for a legitimate authorization server, and a possible connection to a server you do not fully trust. The Python advisory rules out MCP servers built with the SDK, stdio clients, and clients that attach their own tokens or headers.</p>
<p>My position: if users can add their own MCP server URL to your product, assume you are exposed. If your client only talks to one server you run, your risk is that server being compromised.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-do-now">What to do now<a href="https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer#what-to-do-now" class="hash-link" aria-label="Direct link to What to do now" title="Direct link to What to do now" translate="no">​</a></h2>
<ol>
<li class=""><strong>Find it.</strong> Check lockfiles, not just direct requirements, for <code>mcp</code>, <code>@modelcontextprotocol/sdk</code> and <code>@modelcontextprotocol/client</code>.</li>
<li class=""><strong>Upgrade.</strong> Python to <code>mcp</code> 1.30.0 or 2.2.0. TypeScript to <code>@modelcontextprotocol/sdk</code> 1.31.0 or <code>@modelcontextprotocol/client</code> 2.2.0, plus <code>@modelcontextprotocol/core</code> 2.2.0 if you import it directly.</li>
<li class=""><strong>Pin the issuer on machine-to-machine providers.</strong> In Python, pass <code>issuer="https://auth.example.com"</code> to <code>ClientCredentialsOAuthProvider</code> and <code>PrivateKeyJWTOAuthProvider</code>, and move off the deprecated <code>RFC7523OAuthClientProvider</code>, which has no such option. Omitting it emits a deprecation warning, a <code>DeprecationWarning</code> hidden by default on 1.30.0, and it becomes required in 3.0. In TypeScript, pass <code>expectedIssuer: 'https://auth.example.com'</code> to <code>ClientCredentialsProvider</code>, <code>PrivateKeyJwtProvider</code>, <code>StaticPrivateKeyJwtProvider</code> and <code>CrossAppAccessProvider</code>.</li>
<li class=""><strong>Fix what is already stored.</strong> Credentials saved before the upgrade stay unbound. In Python, clear stored client information once so the client registers again, or set <code>issuer</code> on records you pre-registered. In TypeScript, add <code>issuer</code> to saved credentials or clear them. A custom <code>OAuthClientProvider</code> must persist exactly what <code>saveTokens()</code> and <code>saveClientInformation()</code> receive, and, per the <a href="https://github.com/modelcontextprotocol/typescript-sdk/releases/tag/1.31.0" target="_blank" rel="noopener noreferrer" class="">1.31.0 release notes</a>, storage that rejects unknown fields must allow <code>issuer</code>.</li>
<li class=""><strong>Rotate where you cannot rule it out.</strong> If a client may have met an untrusted server, rotate its client secret or signing key and revoke its tokens. Neither advisory publishes indicators of compromise, so I would rotate rather than hunt.</li>
<li class=""><strong>Know what the fix leaves open.</strong> The TypeScript fix does not cover new interactive sign-ins, direct <code>refreshAuthorization()</code> and <code>exchangeAuthorization()</code> calls, or 2.x with <code>skipIssuerMetadataValidation: true</code>. Remove that flag, and complete sign-ins only for servers you trust.</li>
</ol>
<p>If you cannot upgrade yet, connect OAuth clients only to servers you trust. TypeScript 2.0.0 and 2.1.0 can already use <code>expectedIssuer</code>.</p>
<p>After upgrading, Python raises <code>OAuthFlowError</code> with "Authorization server metadata issuer mismatch", and in <a href="https://github.com/modelcontextprotocol/typescript-sdk/releases/tag/v2.2.0" target="_blank" rel="noopener noreferrer" class="">TypeScript 2.2.0</a> <code>fetchToken()</code> throws <code>AuthorizationServerMismatchError</code> before sending anything. Investigate that server.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="an-mcp-server-is-not-an-identity-authority">An MCP server is not an identity authority<a href="https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer#an-mcp-server-is-not-an-identity-authority" class="hash-link" aria-label="Direct link to An MCP server is not an identity authority" title="Direct link to An MCP server is not an identity authority" translate="no">​</a></h2>
<p>In one client, this bug is one team's bad week. In the SDK's OAuth client, it is the default for every client built on it, including teams who chose the SDK so they would never have to think about this code.</p>
<p>The design error is about <a class="" href="https://docs.dwaar.ai/blog/why-agentic-identity-is-the-missing-layer-in-llm-security">identity between agents and tools</a>. The MCP server is where a credential gets used. It should have no say over where that credential gets sent. A long-lived agent secret already has a home, and the client holding it has to know that home itself.</p>
<p>A version check will not close this one. An SCA finding clears when the version moves, while an unpinned provider and an unbound stored credential remain. Treat <code>issuer</code> and <code>expectedIssuer</code> as configuration you audit, and add them to your <a class="" href="https://docs.dwaar.ai/blog/mcp-security-checklist-for-agent-builders">MCP security checklist</a>.</p>
<p>Read release notes for auth changes too. The Python change opened as <a href="https://github.com/modelcontextprotocol/python-sdk/pull/3398" target="_blank" rel="noopener noreferrer" class="">a public hardening PR</a> on 2026-08-26 and <a href="https://github.com/modelcontextprotocol/python-sdk/releases/tag/v1.30.0" target="_blank" rel="noopener noreferrer" class="">shipped on 2026-09-07</a> described as a behaviour change. The vulnerability label came three weeks later.</p>
<p>Today, open the code that constructs your MCP OAuth provider. Write down which authorization server each credential belongs to, then check that the code enforces it rather than asking the server.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/mcp-sdk-oauth-clients-send-credentials-to-a-server-chosen-issuer#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://github.com/modelcontextprotocol/python-sdk/security/advisories/GHSA-qx49-fqc8-xw99" target="_blank" rel="noopener noreferrer" class="">Python advisory GHSA-qx49-fqc8-xw99</a>, Model Context Protocol, 2026-09-28</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/security/advisories/GHSA-6qxp-vccf-f47h" target="_blank" rel="noopener noreferrer" class="">TypeScript advisory GHSA-6qxp-vccf-f47h</a>, Model Context Protocol, 2026-09-30</li>
<li class=""><a href="https://github.com/modelcontextprotocol/python-sdk/releases/tag/v1.30.0" target="_blank" rel="noopener noreferrer" class="">python-sdk v1.30.0 release</a>, Model Context Protocol, 2026-09-07</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/releases/tag/v2.2.0" target="_blank" rel="noopener noreferrer" class="">typescript-sdk v2.2.0 release</a>, Model Context Protocol, 2026-09-28</li>
<li class=""><a href="https://github.com/modelcontextprotocol/python-sdk/pull/3398" target="_blank" rel="noopener noreferrer" class="">python-sdk PR #3398</a>, Model Context Protocol, 2026-08-26</li>
<li class=""><a href="https://github.com/modelcontextprotocol/python-sdk/pull/3431" target="_blank" rel="noopener noreferrer" class="">python-sdk PR #3431, v1.x backport</a>, Model Context Protocol, 2026-09-04</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2887" target="_blank" rel="noopener noreferrer" class="">typescript-sdk PR #2887</a>, Model Context Protocol, 2026-09-28</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/releases/tag/1.31.0" target="_blank" rel="noopener noreferrer" class="">typescript-sdk 1.31.0 release</a>, Model Context Protocol, 2026-09-28</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2888" target="_blank" rel="noopener noreferrer" class="">typescript-sdk PR #2888, v1.x binding</a>, Model Context Protocol, 2026-09-28</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/blob/1.31.0/src/client/auth.ts" target="_blank" rel="noopener noreferrer" class="">typescript-sdk 1.31.0 client source</a>, Model Context Protocol, 2026-09-28</li>
<li class=""><a href="https://github.com/modelcontextprotocol/typescript-sdk/blob/v2.2.0/packages/client/src/client/auth.ts" target="_blank" rel="noopener noreferrer" class="">typescript-sdk v2.2.0 client source</a>, Model Context Protocol, 2026-09-28</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>AI agents</category>
            <category>MCP</category>
            <category>Identity</category>
        </item>
        <item>
            <title><![CDATA[CVE-2026-94545: One Satori Escaping Bug, Moderate in the Library and Critical in next/og]]></title>
            <link>https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og</link>
            <guid>https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og</guid>
            <pubDate>Thu, 24 Sep 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[One Satori SVG escaping flaw, scored Moderate 5.3 by Satori and Critical 9.5 by Next.js. Fixed in satori 0.33.5 and next 16.3.6. Only some apps can reach the critical path.]]></description>
            <content:encoded><![CDATA[<p>Vercel published two advisories for one escaping bug on Tuesday, and they do not agree on how bad it is. <a href="https://github.com/vercel/satori/security/advisories/GHSA-wx4j-mvgx-mqwp" target="_blank" rel="noopener noreferrer" class="">Satori's GHSA-wx4j-mvgx-mqwp</a> calls it Moderate, CVSS 5.3. <a href="https://github.com/vercel/next.js/security/advisories/GHSA-vcvr-r3jv-pc5j" target="_blank" rel="noopener noreferrer" class="">Next.js's GHSA-vcvr-r3jv-pc5j</a> calls it Critical, CVSS 9.5, with remote code execution in the title. Both list CVE-2026-94545, but <a href="https://nextjs.org/blog/nextjs-security-update-september-22-2026" target="_blank" rel="noopener noreferrer" class="">Vercel's security update</a> and the <a href="https://github.com/vercel/next.js/releases/tag/v16.3.6" target="_blank" rel="noopener noreferrer" class="">16.3.6 release notes</a> use only the GHSA IDs, so search for those as well. The question that matters is which advisory describes your app.</p>

































<table><thead><tr><th>Item</th><th>Detail</th></tr></thead><tbody><tr><td>Identifier</td><td>GHSA-wx4j-mvgx-mqwp for Satori, GHSA-vcvr-r3jv-pc5j for Next.js. Both repository advisories list CVE-2026-94545</td></tr><tr><td>Affected</td><td><code>satori</code> <code>&gt;= 0.0.27, &lt; 0.33.5</code>. <code>next</code> <code>&gt;= 16.2.0, &lt; 16.3.6</code>, Node.js <code>ImageResponse</code> in <code>next/og</code> only</td></tr><tr><td>Fixed in</td><td><code>satori</code> 0.33.5, <code>next</code> 16.3.6</td></tr><tr><td>Severity</td><td>Vendor scored, CVSS v4: Moderate 5.3 in the Satori advisory, Critical 9.5 in the Next.js advisory</td></tr><tr><td>Exploited as of 2026-09-24</td><td>No exploitation or public proof of concept reported in either advisory. Not listed in <a href="https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json" target="_blank" rel="noopener noreferrer" class="">CISA KEV</a></td></tr><tr><td>Disclosed</td><td>2026-09-22</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-is-broken">What is broken<a href="https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og#what-is-broken" class="hash-link" aria-label="Direct link to What is broken" title="Direct link to What is broken" translate="no">​</a></h2>
<p>Satori says it <a href="https://github.com/vercel/satori/security/advisories/GHSA-wx4j-mvgx-mqwp" target="_blank" rel="noopener noreferrer" class="">does not properly escape certain values</a> before including them in generated SVG. The useful detail is in <a href="https://github.com/vercel/satori/commit/26a52affc031216fee5882b6e965c8dbc7ac1782" target="_blank" rel="noopener noreferrer" class="">the fix commit</a> from <a href="https://github.com/vercel/satori/pull/814" target="_blank" rel="noopener noreferrer" class="">PR #814</a>.</p>
<p>Before the fix, Satori assembled much of its SVG by string interpolation, in several places:</p>
<ul>
<li class="">The shared <code>buildXMLString</code> helper in <code>src/utils.ts</code> wrote attribute values straight into markup.</li>
<li class="">The inline <code>&lt;svg&gt;</code> path in <code>src/handler/preprocess.ts</code> interpolated attribute values, style values and element names, and emitted non-object children with <code>String(node)</code>.</li>
<li class=""><code>rect.ts</code>, <code>text.ts</code> and others built wrapper tags such as <code>&lt;g clip-path="url(#...)"&gt;</code> by hand.</li>
<li class=""><code>src/handler/expand.ts</code> copied caller-supplied style keys starting with <code>_</code> into the serialized style. Those are Satori's internal fields, such as <code>_inheritedClipPathId</code>, and they are later written into markup.</li>
</ul>
<p>So a color or font-family containing a double quote could close the attribute and add new SVG elements. The regression tests added with the fix do exactly this, with a color value that injects a <code>path</code> element.</p>
<p>The core of it, verbatim from the commit:</p>
<div class="language-diff codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-diff codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token plain">-      attrString += ` ${k}="${_v}"`</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">+      attrString += ` ${k}="${escapeXMLAttribute(_v)}"`</span><br></div></code></pre></div></div>
<p>The rest of the fix is the part I would copy. <code>buildXMLString</code> now validates element and attribute names and throws on bad ones, because, as the code comment puts it, "Names cannot be escaped". Text children get escaped, the hand-built wrapper tags now go through the helper, and underscore-prefixed style keys from callers throw <code>Invalid style property</code>. <code>&amp;</code> is now also encoded in embedded SVG data URLs so that inner XML entities survive the enclosing parse. One serializer everything goes through beats many call sites that each have to remember.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-vercel-has-and-has-not-said-about-the-rce-step">What Vercel has and has not said about the RCE step<a href="https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og#what-vercel-has-and-has-not-said-about-the-rce-step" class="hash-link" aria-label="Direct link to What Vercel has and has not said about the RCE step" title="Direct link to What Vercel has and has not said about the RCE step" translate="no">​</a></h2>
<p>That is markup injection. The Next.js advisory says the Node.js <code>ImageResponse</code> "is affected by an upstream vulnerability. This can lead to remote code execution." Affected apps, it says, "pass attacker-controlled values into SVG content, attributes, or styles during image generation". Its example is a query parameter inside an SVG <code>&lt;title&gt;</code>.</p>
<p>Neither advisory explains the step. <a href="https://nextjs.org/blog/nextjs-security-update-september-22-2026" target="_blank" rel="noopener noreferrer" class="">Vercel's security update</a> says that, under specific conditions, improper escaping in Satori's SVG output could lead to RCE "due to vulnerabilities in other upstream dependencies", and that the fix upgrades them. It does not name those dependencies or the path. I am not going to guess. Treat the RCE as the maintainer's stated assessment, with those specifics unpublished, and act on it anyway.</p>
<p>The vectors tell the rest. Both share <code>AV:N/AC:L/AT:P/PR:N</code>. <a href="https://github.com/vercel/satori/security/advisories/GHSA-wx4j-mvgx-mqwp" target="_blank" rel="noopener noreferrer" class="">Satori's</a> has passive user interaction and no impact on the vulnerable system, with high confidentiality and integrity impact on subsequent systems. <a href="https://github.com/vercel/next.js/security/advisories/GHSA-vcvr-r3jv-pc5j" target="_blank" rel="noopener noreferrer" class="">Next.js's</a> has no user interaction and high impact across all six metrics. Satori says it plainly: "The impact depends on how the generated SVG is consumed."</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="same-package-very-different-risk">Same package, very different risk<a href="https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og#same-package-very-different-risk" class="hash-link" aria-label="Direct link to Same package, very different risk" title="Direct link to Same package, very different risk" translate="no">​</a></h2>
<p>Take two apps on <code>next</code> 16.3.5. One renders social cards from strings fixed at build time. The other reads a title from the query string. Version matching reports the same finding for both, and by the Next.js advisory only the second is exposed: apps that never pass attacker-controlled values into SVG are not affected, and neither is the Edge <code>ImageResponse</code>.</p>
<div class="language-tsx codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-tsx codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token comment">// Illustration, not from the advisory: the shape to search for.</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token keyword">const</span><span class="token plain"> title </span><span class="token operator">=</span><span class="token plain"> </span><span class="token keyword">new</span><span class="token plain"> </span><span class="token class-name constant">URL</span><span class="token punctuation">(</span><span class="token plain">request</span><span class="token punctuation">.</span><span class="token property-access">url</span><span class="token punctuation">)</span><span class="token punctuation">.</span><span class="token property-access">searchParams</span><span class="token punctuation">.</span><span class="token method function property-access">get</span><span class="token punctuation">(</span><span class="token string">'title'</span><span class="token punctuation">)</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token keyword">return</span><span class="token plain"> </span><span class="token keyword">new</span><span class="token plain"> </span><span class="token class-name">ImageResponse</span><span class="token punctuation">(</span><span class="token tag punctuation">&lt;</span><span class="token tag">svg</span><span class="token tag punctuation">&gt;</span><span class="token plain-text">&lt;title&gt;</span><span class="token punctuation">{</span><span class="token plain">title</span><span class="token punctuation">}</span><span class="token tag punctuation">&lt;/</span><span class="token tag">title</span><span class="token tag punctuation">&gt;</span><span class="token tag punctuation">&lt;/</span><span class="token tag">svg</span><span class="token tag punctuation">&gt;</span><span class="token punctuation">)</span><br></div></code></pre></div></div>
<p>My position has three parts.</p>
<p>Use reachability to order the queue, not to skip the patch. Your read of a data flow can be wrong, and the fix is a version bump. <a class="" href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof">A finding is a claim</a>, and a version match only starts the investigation.</p>
<p>Read "attacker-controlled" broadly. A query parameter is the obvious case. A display name, a product title or a CMS field that outside users can edit is the same input with a delay.</p>
<p>Do not treat 5.3 as a ceiling if you call Satori directly. That score describes a library whose risk lives downstream, and the Next.js advisory shows a consumer can change the answer completely. <a class="" href="https://docs.dwaar.ai/blog/cvss-is-not-risk-how-to-use-epss-and-kev">CVSS is not risk</a>, and this CVE has two scores to prove it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-check-now">What to check now<a href="https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og#what-to-check-now" class="hash-link" aria-label="Direct link to What to check now" title="Direct link to What to check now" translate="no">​</a></h2>
<ol>
<li class=""><strong>Check your <code>next</code> version.</strong> If it is <code>&gt;= 16.2.0, &lt; 16.3.6</code>, go to <a href="https://github.com/vercel/next.js/releases/tag/v16.3.6" target="_blank" rel="noopener noreferrer" class="">16.3.6</a> or later. Vercel also shipped <a href="https://github.com/vercel/next.js/releases/tag/v15.5.26" target="_blank" rel="noopener noreferrer" class="">15.5.26</a>, but <a href="https://nextjs.org/blog/nextjs-security-update-september-22-2026" target="_blank" rel="noopener noreferrer" class="">its security update</a> labels that release hardening only and says Next.js 15.x is not affected by the RCE. Do not rely on searching your lockfile for <code>satori</code>: <a href="https://github.com/vercel/next.js/commit/868fad38690d72088868f299fa2bef339b26838e" target="_blank" rel="noopener noreferrer" class="">the Next.js fix</a> patches <code>@vercel/og@0.11.1</code> and updates compiled og bundles, so I would treat the <code>next</code> version as the thing to check.</li>
<li class=""><strong>Check direct Satori use.</strong> Move to 0.33.5 or later, then find out what consumes the SVG you generate.</li>
<li class=""><strong>Find your exposed routes.</strong> Look for <code>ImageResponse</code> from <code>next/og</code> on the Node.js runtime, then trace whether query parameters, headers or body values reach the JSX.</li>
<li class=""><strong>If you cannot upgrade today,</strong> stop passing attacker-controlled values into SVG content, attributes or styles. This is a stopgap. Satori says "No complete workaround exists besides upgrading."</li>
<li class=""><strong>Expect new errors.</strong> The Satori fix throws <code>Invalid XML element name</code>, <code>Invalid XML attribute name</code> and <code>Invalid style property</code>. Code that deliberately passes underscore-prefixed styles will now fail, so test image routes before you ship.</li>
<li class=""><strong>Look at your logs, with low expectations.</strong> The sources publish no detection guidance or indicators. My inference from the fix: look for quotes or angle brackets followed by <code>/&gt;</code> or tag names in parameters feeding image routes and, after upgrading, the error strings above, which come from the Satori diff. Finding nothing proves nothing.</li>
</ol>
<p>Both advisories credit RaghavMaheshwari124 and rafabd1 as finders.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="one-check-to-run-this-week">One check to run this week<a href="https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og#one-check-to-run-this-week" class="hash-link" aria-label="Direct link to One check to run this week" title="Direct link to One check to run this week" translate="no">​</a></h2>
<p>The reachability question can be answered by hand. Open your busiest Next.js repository, list every route that returns <code>ImageResponse</code>, and write down where each rendered string comes from. Any string you cannot trace to something you control goes to the top of the upgrade queue.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/satori-svg-escaping-cve-2026-94545-rce-in-next-og#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://github.com/vercel/satori/security/advisories/GHSA-wx4j-mvgx-mqwp" target="_blank" rel="noopener noreferrer" class="">Improper escaping in Satori-generated SVG (GHSA-wx4j-mvgx-mqwp)</a>, Vercel via GitHub Security Advisories, 2026-09-22</li>
<li class=""><a href="https://github.com/vercel/next.js/security/advisories/GHSA-vcvr-r3jv-pc5j" target="_blank" rel="noopener noreferrer" class="">Remote Code Execution in next/og ImageResponse (GHSA-vcvr-r3jv-pc5j)</a>, Vercel via GitHub Security Advisories, 2026-09-22</li>
<li class=""><a href="https://nextjs.org/blog/nextjs-security-update-september-22-2026" target="_blank" rel="noopener noreferrer" class="">Next.js Security Update for a Critical Upstream Issue</a>, Next.js blog, 2026-09-22</li>
<li class="">Repository advisory records showing CVE-2026-94545, GitHub REST API: <a href="https://api.github.com/repos/vercel/next.js/security-advisories/GHSA-vcvr-r3jv-pc5j" target="_blank" rel="noopener noreferrer" class="">Next.js</a>, last updated 2026-09-22, and <a href="https://api.github.com/repos/vercel/satori/security-advisories/GHSA-wx4j-mvgx-mqwp" target="_blank" rel="noopener noreferrer" class="">Satori</a>, last updated 2026-09-22</li>
<li class=""><a href="https://github.com/vercel/satori/pull/814" target="_blank" rel="noopener noreferrer" class="">fix: Harden SVG serialization, PR #814</a>, Vercel (satori repository), 2026-09-22</li>
<li class=""><a href="https://github.com/vercel/satori/commit/26a52affc031216fee5882b6e965c8dbc7ac1782" target="_blank" rel="noopener noreferrer" class="">fix: Harden SVG serialization (#814), commit 26a52af</a>, Vercel (satori repository), 2026-09-22</li>
<li class=""><a href="https://github.com/vercel/next.js/commit/868fad38690d72088868f299fa2bef339b26838e" target="_blank" rel="noopener noreferrer" class="">Harden next/og SVG serialization, commit 868fad3</a>, Vercel (next.js repository), 2026-09-22</li>
<li class=""><a href="https://github.com/vercel/next.js/releases/tag/v16.3.6" target="_blank" rel="noopener noreferrer" class="">Next.js v16.3.6 release</a>, Vercel via GitHub releases, 2026-09-22</li>
<li class=""><a href="https://github.com/vercel/next.js/releases/tag/v15.5.26" target="_blank" rel="noopener noreferrer" class="">Next.js v15.5.26 release</a>, Vercel via GitHub releases, 2026-09-22</li>
<li class=""><a href="https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json" target="_blank" rel="noopener noreferrer" class="">Known Exploited Vulnerabilities catalog</a>, CISA, continuously updated feed with no single publication date</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>Web frameworks</category>
            <category>Open source</category>
            <category>Reachability</category>
        </item>
        <item>
            <title><![CDATA[The APIs Nobody Told You About]]></title>
            <link>https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about</link>
            <guid>https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about</guid>
            <pubDate>Wed, 16 Sep 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[A spec describes the API you meant to ship. Why OWASP's API9 inventory gaps and API1 broken object level authorization stay invisible when you only match known CVEs.]]></description>
            <content:encoded><![CDATA[<p>When someone asks how many APIs you run, the first thing you reach for is the OpenAPI spec. That file is a statement of intent. It records what the team meant to ship. What answers on a port tonight is a different question, and whether that thing checks who is asking is a third.</p>
<p>The usual program is sensible: keep the spec current, point a scanner of known CVEs at the hosts it names, patch what comes back. It fails in two places. The scan covers the hosts you handed it, and those came from the list that is wrong. And a CVE scanner compares component versions against known vulnerabilities, so it can tell you the gateway in front of your API has a published flaw, but not whether user B may read user A's invoice. That is my reasoning from the definitions, not a measured result, and narrower than <a class="" href="https://docs.dwaar.ai/blog/why-scanning-alone-is-not-a-security-program">Why Scanning Alone Is Not a Security Program</a>.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-an-inventory-has-to-answer">What an inventory has to answer<a href="https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about#what-an-inventory-has-to-answer" class="hash-link" aria-label="Direct link to What an inventory has to answer" title="Direct link to What an inventory has to answer" translate="no">​</a></h2>
<p>The <a href="https://owasp.org/API-Security/editions/2023/en/0x11-t10/" target="_blank" rel="noopener noreferrer" class="">OWASP API Security Top 10, 2023 edition</a>, names the first gap at API9, Improper Inventory Management. Per the <a href="https://owasp.org/API-Security/editions/2023/en/0xa9-improper-inventory-management/" target="_blank" rel="noopener noreferrer" class="">API9 page</a>, you have a documentation blindspot when the purpose of a host is unclear and nobody can say which environment it runs in (production, staging, test, development), who should have network access (public, internal, partners), and which version is running. OWASP adds a data flow blindspot: sensitive data going to a third party with no business justification or no inventory of the flow.</p>
<p>Those are questions about a running system. A spec can assert the answers. It cannot confirm them.</p>
<p>The same page says why this matters. Attackers "usually get unauthorized access through old API versions or endpoints left running unpatched and using weaker security requirements". It adds: "Sometimes different API versions/deployments are connected to the same database with real data." The forgotten host is not necessarily holding forgotten data.</p>
<p>OWASP's first API9 scenario, an illustration and not a reported incident, shows the shape. A beta host runs the same API as production, but rate limiting lives in a separate component that is not in place on beta, so a researcher brute forces a 6-digit reset token and can reset any user's password. The two hosts have the same API. The difference is entirely in what surrounds them.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="every-source-is-a-partial-witness">Every source is a partial witness<a href="https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about#every-source-is-a-partial-witness" class="hash-link" aria-label="Direct link to Every source is a partial witness" title="Direct link to Every source is a partial witness" translate="no">​</a></h2>
<p>No single system holds the true list. This table is my own assessment, not taken from a standard:</p>








































<table><thead><tr><th>Evidence</th><th>What it tells you</th><th>What it cannot tell you</th></tr></thead><tbody><tr><td>Hand-written spec</td><td>What the team intended</td><td>Anything shipped outside that intent, or after the file went stale</td></tr><tr><td>Spec generated in CI/CD</td><td>What the built service declares</td><td>Older deployments still running, hosts outside that pipeline</td></tr><tr><td>Gateway configuration</td><td>Routes that pass through the gateway</td><td>Hosts that do not</td></tr><tr><td>Traffic logs</td><td>What is being called</td><td>Routes nobody called in the window</td></tr><tr><td>Source code</td><td>Which handlers exist, whether a check is written</td><td>What is deployed, and who can reach it</td></tr><tr><td>Known CVE scan</td><td>Known flaws in components on hosts you named</td><td>Hosts nobody named, and who may read what</td></tr></tbody></table>
<p>OWASP recommends generating documentation automatically in CI/CD, and if you change one thing about your spec, change that. But the value is in the disagreement between the rows. A path in the logs that the spec has never heard of, a route in code with no gateway entry: those are the APIs nobody told you about. <a class="" href="https://docs.dwaar.ai/blog/you-cannot-secure-what-you-cannot-list">You Cannot Secure What You Cannot List</a> makes the wider case.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-bug-with-no-version-number">The bug with no version number<a href="https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about#the-bug-with-no-version-number" class="hash-link" aria-label="Direct link to The bug with no version number" title="Direct link to The bug with no version number" translate="no">​</a></h2>
<p>Now suppose your list is perfect. The top entry on OWASP's list is still waiting.</p>
<p><a href="https://owasp.org/API-Security/editions/2023/en/0xa1-broken-object-level-authorization/" target="_blank" rel="noopener noreferrer" class="">API1:2023, Broken Object Level Authorization</a>, carries OWASP's ratings of Widespread for prevalence and Easy for exploitability. Those are ratings, not measured rates. The rule: "Every API endpoint that receives an ID of an object, and performs any action on the object, should implement object-level authorization checks."</p>
<p>This sentence explains why it evades the usual tooling: "it's by design that the user will have access to the vulnerable API endpoint/function. The violation happens at the object level, by manipulating the ID." <a href="https://cwe.mitre.org/data/published/cwe_v4.20.pdf#page=1412" target="_blank" rel="noopener noreferrer" class="">CWE-639</a>, one of the two weaknesses OWASP maps API1 to, adds that the attacker is likely an authenticated user, and notes that BOLA appears to overlap with it but may apply to other weaknesses too.</p>
<p>So the request is unremarkable: valid token, permitted route, well formed ID. The flaw belongs to a handler your team wrote, so a scanner matching known CVEs against component versions has no package, version or advisory to match. That is my inference from these definitions, not a measurement.</p>
<p>OWASP's BOLA scenario describes an e-commerce platform with a pattern like <code>/shops/{shopName}/revenue_data.json</code>. An attacker iterates over shop names and reads the sales data of thousands of stores. OWASP also warns that comparing the session's user ID with the ID parameter "isn't a sufficient solution", and as I read it this scenario shows why. A shop name is not a user ID. Whether this caller may see this shop's revenue depends on your data model, so the answer can only live in your code and your tests.</p>
<p>Here the two problems join up. "Every API endpoint" is a claim about a set, and you cannot test a set you have not enumerated. Inventory does not fix authorization. It defines how much authorization work there is.</p>
<p>One API can carry a lot. T-Mobile's <a href="https://www.sec.gov/Archives/edgar/data/1283699/000119312523010949/d641142d8k.htm" target="_blank" rel="noopener noreferrer" class="">Form 8-K of 19 January 2023</a> says it identified on 5 January 2023 that "a bad actor was obtaining data through a single Application Programming Interface". Its preliminary result was approximately 37 million current postpaid and prepaid customer accounts, many of which did not include the full data set. The filing does not call the flaw BOLA, so I will not either, and that number counts accounts, not harm.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-i-am-not-arguing">What I am not arguing<a href="https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about#what-i-am-not-arguing" class="hash-link" aria-label="Direct link to What I am not arguing" title="Direct link to What I am not arguing" translate="no">​</a></h2>
<p>CVE scanning stays. It finds vulnerable components, gateways included, on the APIs you know about. Specs stay too. My complaint is with hand-maintained ones treated as the inventory.</p>
<p>I do not think finding BOLA needs a new category of tool. Alongside a proper authorization mechanism, OWASP's prevention advice includes tests: "Write tests to evaluate the vulnerability of the authorization mechanism. Do not deploy changes that make the tests fail."</p>
<p>And I am giving you no figure for how many organizations run APIs they do not know about, because I could not find a neutral one.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="two-accounts-one-route">Two accounts, one route<a href="https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about#two-accounts-one-route" class="hash-link" aria-label="Direct link to Two accounts, one route" title="Direct link to Two accounts, one route" translate="no">​</a></h2>
<p>Pick one production service. Diff the routes in its spec against the distinct method and path pairs in your gateway or load balancer logs. For anything in the logs and not in the spec, write down environment, audience and version. If nobody can answer, that is your first inventory entry.</p>
<p>Then take one route that accepts an object ID and create two test identities that own different objects.</p>
<div class="language-text codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-text codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token plain"># Illustration written for this post, not from any source.</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">GET /invoices/{id-owned-by-A}  with A's token   expect success</span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">GET /invoices/{id-owned-by-A}  with B's token   expect refusal</span><br></div></code></pre></div></div>
<p>If the second request succeeds, you have a bug no version bump will fix. If it is refused, commit the pair as a test that blocks deployment, and move to the next route on the list you just built.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/the-apis-nobody-told-you-about#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">OWASP API Security Project, <a href="https://owasp.org/API-Security/editions/2023/en/0x11-t10/" target="_blank" rel="noopener noreferrer" class="">OWASP Top 10 API Security Risks, 2023</a>, 2023 edition</li>
<li class="">OWASP API Security Project, <a href="https://owasp.org/API-Security/editions/2023/en/0xa1-broken-object-level-authorization/" target="_blank" rel="noopener noreferrer" class="">API1:2023 Broken Object Level Authorization</a>, 2023 edition</li>
<li class="">OWASP API Security Project, <a href="https://owasp.org/API-Security/editions/2023/en/0xa9-improper-inventory-management/" target="_blank" rel="noopener noreferrer" class="">API9:2023 Improper Inventory Management</a>, 2023 edition</li>
<li class="">MITRE, <a href="https://cwe.mitre.org/data/published/cwe_v4.20.pdf#page=1412" target="_blank" rel="noopener noreferrer" class="">CWE-639: Authorization Bypass Through User-Controlled Key</a>, CWE 4.20, 30 April 2026, PDF page 1412</li>
<li class="">T-Mobile US, Inc., via SEC, <a href="https://www.sec.gov/Archives/edgar/data/1283699/000119312523010949/d641142d8k.htm" target="_blank" rel="noopener noreferrer" class="">Form 8-K</a>, 19 January 2023</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>API security</category>
            <category>Inventory</category>
        </item>
        <item>
            <title><![CDATA[The EU Cyber Resilience Act Reporting Clock Starts This Week]]></title>
            <link>https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts</link>
            <guid>https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts</guid>
            <pubDate>Wed, 09 Sep 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Article 14 of the CRA applies from 11 September 2026, with a 24-hour early warning for actively exploited vulnerabilities. What it asks of manufacturers, and what filing on time does not prove.]]></description>
            <content:encoded><![CDATA[<p>On Friday a new duty lands on manufacturers with a software or hardware product on the EU market. If a vulnerability in that product is being actively exploited, an early warning is due within 24 hours of becoming aware of it. Deciding who files that warning is the easy part. The harder part comes first: when a flaw in a component is reported as exploited, showing whether that component is in what you shipped, and whether the flaw can be or has been exploited there. That is where I would spend this week.</p>
<p><a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847" target="_blank" rel="noopener noreferrer" class="">Article 71 of the Cyber Resilience Act</a> applies Article 14 from 11 September 2026 and the main obligations from 11 December 2027 (Chapter IV, on notifying conformity assessment bodies, applied earlier, from 11 June 2026), so reporting arrives well before the design requirements, CE marking and SBOM duty. <a href="https://digital-strategy.ec.europa.eu/en/policies/cra-summary" target="_blank" rel="noopener noreferrer" class="">The Commission</a> says it reaches "all products with digital elements that have been made available on the Union market", including those already out there.</p>
<p>This is an engineer's reading, not legal advice. Article numbers refer to the <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847" target="_blank" rel="noopener noreferrer" class="">regulation on EUR-Lex</a>. Paragraph numbers refer to the <a href="https://ec.europa.eu/newsroom/dae/redirection/document/131456" target="_blank" rel="noopener noreferrer" class="">Commission guidance dated 27 July 2026</a>, which says it is not binding and that an authoritative interpretation can only come from the Court of Justice (para 8).</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-the-clock-asks">What the clock asks<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#what-the-clock-asks" class="hash-link" aria-label="Direct link to What the clock asks" title="Direct link to What the clock asks" translate="no">​</a></h2>

























<table><thead><tr><th>Step, for an actively exploited vulnerability</th><th>Deadline</th><th>Runs from</th></tr></thead><tbody><tr><td>Early warning</td><td>24 hours</td><td>Becoming aware of it</td></tr><tr><td>Vulnerability notification</td><td>72 hours</td><td>Becoming aware of it</td></tr><tr><td>Final report</td><td>14 days</td><td>A corrective or mitigating measure being available</td></tr></tbody></table>
<p>Reports go simultaneously to the CSIRT designated as coordinator and to ENISA, through the single reporting platform of Article 16. Article 14(8) adds a duty to inform impacted users. Failing Articles 13 and 14 can draw <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847" target="_blank" rel="noopener noreferrer" class="">fines of up to EUR 15 million or 2.5 percent of worldwide annual turnover</a>, whichever is higher.</p>
<p>The trigger is narrower than it sounds. Article 3(42) requires "reliable evidence that a malicious actor has exploited it in a system without permission of the system owner". By that definition, a high severity score with no evidence of exploitation does not start this clock. Severe incidents, defined in Article 14(5), are a second track, so read the article itself for their final report.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-obvious-plan-and-where-it-breaks">The obvious plan, and where it breaks<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#the-obvious-plan-and-where-it-breaks" class="hash-link" aria-label="Direct link to The obvious plan, and where it breaks" title="Direct link to The obvious plan, and where it breaks" translate="no">​</a></h2>
<p>The obvious response is to treat this as a filing problem: name an owner, write a runbook, rehearse the submission. Do all of that. It still breaks at the first step, because a runbook starts once someone says "this is in our product", and nothing in the runbook produces that sentence.</p>
<p>A runbook fits a researcher emailing you about your own code. It fits badly when the flaw sits in a library three layers down, and the <a href="https://ec.europa.eu/newsroom/dae/redirection/document/131456" target="_blank" rel="noopener noreferrer" class="">Commission guidance</a> draws a line there. A flaw of third-party origin has to be notified when your product contains it and it is actively exploited there. If it "cannot be exploited" in your product, for example "because the vulnerable code is not reachable", or "has not been exploited" in your product, it "is not subject to mandatory reporting" for you (para 218). You can still notify voluntarily under Article 15, and the same paragraph points to the vulnerability handling requirements and an upstream report to the component's maintainer.</p>
<p>That narrows the duty without shrinking the work. You cannot show that a flaw is unreachable or unexploited in a product until you know the component is there. So the inventory question comes before the legal one. That ordering is my inference, not the guidance's wording.</p>
<p>Log4j shows the order. <a href="https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json" target="_blank" rel="noopener noreferrer" class="">CISA added CVE-2021-44228 to its Known Exploited Vulnerabilities catalog</a> on 10 December 2021. Article 14 was years away, and on the Commission's reading a listing like that would not by itself make a flaw reportable by a manufacturer. As I read that episode, the difficulty was never understanding the bug. It was answering which of my products contain this component, and no later question can start until that one is answered.</p>
<p>It is tempting to read "becoming aware" as an exit: no looking, no clock. I would not build on that. Para 213 covers the case where a customer, an authority, a media organisation or another source brings a potential vulnerability to your attention. The manufacturer "should assess the suspicious event immediately", and is aware "when, after such an initial assessment, it has a reasonable degree of certainty" that a vulnerability contained in its product is being actively exploited. Para 214 adds that "the emphasis should be on prompt action to carry out the initial assessment". An assessment that stalls because nobody knows where the component is does not look like prompt action to me. Take the edge cases to your counsel. And from December 2027, Article 13(5) requires due diligence on third-party components, including open source.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-you-have-to-be-able-to-answer">What you have to be able to answer<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#what-you-have-to-be-able-to-answer" class="hash-link" aria-label="Direct link to What you have to be able to answer" title="Direct link to What you have to be able to answer" translate="no">​</a></h2>
<p><a class="" href="https://docs.dwaar.ai/blog/you-cannot-secure-what-you-cannot-list">You cannot secure what you cannot list</a>, and you cannot run that initial assessment on it either. That is my inference, not the text. Article 14 does not mandate an SBOM. Article 3(39) defines one, and the requirement itself sits in Annex I and applies from December 2027. My position is to build it now anyway. It is only one input, and <a class="" href="https://docs.dwaar.ai/blog/what-an-sbom-tells-you-and-what-it-does-not">it has limits</a>. The clock needs four answers:</p>






























<table><thead><tr><th>Question</th><th>Why the clock needs it</th><th>Where I would get it</th></tr></thead><tbody><tr><td>Which products contain this component?</td><td>First filter: without it there is nothing to assess, but containing the component does not by itself trigger reporting</td><td>SBOMs per release, plus repositories and registries</td></tr><tr><td>Which versions are on the market?</td><td>Reporting reaches products already available</td><td>Release and support records</td></tr><tr><td>Who runs them?</td><td>Article 14(8) points at impacted users</td><td>Customer and update records</td></tr><tr><td>Has it been exploited in your product?</td><td>The Commission's non-binding guidance (para 218) reads the duty as covering exploitation in your product, so KEV alone does not settle it</td><td>Reachability analysis, your logs, reports from customers and third parties</td></tr></tbody></table>
<p>KEV is the cheap starting signal. Its feed holds roughly 1,700 entries dated up to this week, and intersecting it with your component records is mechanical. A match tells you CISA lists the flaw as exploited. It does not tell you the flaw was exploited in your product, so treat it as the start of the initial assessment, not the result.</p>
<p>There is a gap to keep in view. Article 3(42) speaks of exploitation "in a system" without saying whose, while Article 14(1) covers a vulnerability "contained in the product". The test of exploitation in your own product is the Commission's reading of those words, and only the Court of Justice can settle it. I would plan around para 218 and write down how you reached each conclusion.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="filing-on-time-is-not-being-secure">Filing on time is not being secure<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#filing-on-time-is-not-being-secure" class="hash-link" aria-label="Direct link to Filing on time is not being secure" title="Direct link to Filing on time is not being secure" translate="no">​</a></h2>
<p>A filing proves a process ran. It does not prove the flaw got fixed, or fixed quickly. Nothing in the first table sets a date for the fix, and the final report timer starts only once a measure exists.</p>
<p>Set that against the summary of findings in <a href="https://www.verizon.com/business/resources/reports/2025-dbir-data-breach-investigations-report.pdf" target="_blank" rel="noopener noreferrer" class="">Verizon's 2025 DBIR</a> (p. 10). Vulnerability exploitation reached 20 percent of breaches as an initial access vector. Edge devices and VPNs went from 3 to 22 percent of the targets of that exploitation, and only about 54 percent of those edge device vulnerabilities were fully remediated throughout the year, at a median of 32 days. A manufacturer can meet every Article 14 deadline while its customers sit inside that window.</p>
<p>The same gap comes with the CE mark. The <a href="https://digital-strategy.ec.europa.eu/en/policies/cra-summary" target="_blank" rel="noopener noreferrer" class="">Commission summary</a> says manufacturers can generally choose a self-assessment (module A), while important or critical categories may require a third party. So the mark does not by itself imply that anyone independent checked.</p>
<p>This is my argument, not a measurement. I have no source comparing CRA compliance with breach outcomes. Defenders of the regulation have a real point too: fast, mandatory, shared reporting is a security gain, since the coordinating CSIRT forwards notifications to CSIRTs in other Member States where the product is available. I accept that, but do not let "filed within 24 hours" become the number your security program reports upward.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-stops-applying">Where this stops applying<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#where-this-stops-applying" class="hash-link" aria-label="Direct link to Where this stops applying" title="Direct link to Where this stops applying" translate="no">​</a></h2>
<p>The duty covers actively exploited vulnerabilities and severe incidents, not your backlog, so it is no vulnerability program. Per the summary, micro and small enterprises may not be fined for missing the 24-hour deadline, and open-source stewards are not fined at all.</p>
<p>Settle scope first. The guidance says software that "executes remotely and is merely accessed by the user" is not, on that basis alone, a product with digital elements (para 21), so a pure SaaS service with no distributed product may fall outside. Remote data processing solutions do count as part of a product, including for Article 14 reporting (para 178), so do not assume. Then put paras 213 and 218 next to your runbook: they set the awareness test and the third-party component rule your triage has to apply.</p>
<p>You can run the drill below with a spreadsheet.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-it-before-friday">Try it before Friday<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#try-it-before-friday" class="hash-link" aria-label="Direct link to Try it before Friday" title="Direct link to Try it before Friday" translate="no">​</a></h2>
<p>Open the <a href="https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json" target="_blank" rel="noopener noreferrer" class="">KEV feed</a> and take the newest entry naming a component you believe you use, or Log4j if none fits. This is a rehearsal: on the guidance's reading, a KEV entry does not by itself start the legal clock. Start a 24-hour timer and answer: which shipped products contain it, whether it can be or has been exploited in them, which versions are on the EU market, which users you would tell, and who files. Count only what records show, not what someone remembers. "None" is valid only if you can show how you know. Every question still open when the timer ends is your CRA work list.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/the-eu-cyber-resilience-act-reporting-clock-starts#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847" target="_blank" rel="noopener noreferrer" class="">Regulation (EU) 2024/2847</a>, EUR-Lex, adopted 23 October 2024, Official Journal 20 November 2024</li>
<li class=""><a href="https://ec.europa.eu/newsroom/dae/redirection/document/131456" target="_blank" rel="noopener noreferrer" class="">Commission guidance on the application of Regulation (EU) 2024/2847 (Cyber Resilience Act), C(2026) 5252 final, annex</a>, European Commission, 27 July 2026</li>
<li class=""><a href="https://digital-strategy.ec.europa.eu/en/policies/cra-summary" target="_blank" rel="noopener noreferrer" class="">The Cyber Resilience Act, summary</a>, European Commission, 3 December 2025</li>
<li class=""><a href="https://digital-strategy.ec.europa.eu/en/policies/cyber-resilience-act" target="_blank" rel="noopener noreferrer" class="">Cyber Resilience Act</a>, European Commission, updated 7 September 2026</li>
<li class=""><a href="https://www.verizon.com/business/resources/reports/2025-dbir-data-breach-investigations-report.pdf" target="_blank" rel="noopener noreferrer" class="">2025 Data Breach Investigations Report</a>, Verizon, 2025, summary of findings, p. 10</li>
<li class=""><a href="https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json" target="_blank" rel="noopener noreferrer" class="">Known Exploited Vulnerabilities catalog (JSON)</a>, CISA, CVE-2021-44228 added 10 December 2021</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>Compliance</category>
            <category>SBOM</category>
        </item>
        <item>
            <title><![CDATA[A Finding Is a Claim. Exploitability Is the Proof.]]></title>
            <link>https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof</link>
            <guid>https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof</guid>
            <pubDate>Wed, 02 Sep 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[A scanner finding says a flaw may apply. A demonstrated exploit shows it works on your build. What counts as proof, how to test with authorization, and why a failed attempt proves little.]]></description>
            <content:encoded><![CDATA[<p>You file a ticket for a critical vulnerability in a library. The service owner replies that the vulnerable function is never called. You reply with the score. They reply with the sprint plan. The ticket stays open, and both of you still think you are right.</p>
<p>Nobody in that thread has evidence. The scanner made a claim: something matched, so this flaw may apply here. The owner made a counterclaim from memory of the code. The argument ends when someone tries the thing and shows what happened.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="better-scoring-does-not-settle-it">Better scoring does not settle it<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#better-scoring-does-not-settle-it" class="hash-link" aria-label="Direct link to Better scoring does not settle it" title="Direct link to Better scoring does not settle it" translate="no">​</a></h2>
<p>The obvious fix is better prioritisation, and it is worth having. <a class="" href="https://docs.dwaar.ai/blog/cvss-is-not-risk-how-to-use-epss-and-kev">CVSS Is Not Risk</a> covers how to use EPSS and KEV. But look at what each signal describes.</p>








































<table><thead><tr><th>Signal</th><th>What it tells you</th><th>What it leaves open</th></tr></thead><tbody><tr><td><a href="https://www.first.org/cvss/v4-0/specification-document" target="_blank" rel="noopener noreferrer" class="">CVSS v4</a> Base score</td><td>Severity relative to other vulnerabilities</td><td>Whether anyone exploits it</td></tr><tr><td>CVSS v4 Exploit Maturity</td><td>Attacks reported, public proof-of-concept code only, or neither known</td><td>Whether it works on your build</td></tr><tr><td><a href="https://certcc.github.io/SSVC/reference/decision_points/exploitation/" target="_blank" rel="noopener noreferrer" class="">SSVC Exploitation</a></td><td>None, Public PoC, or Active with reliable evidence of real attackers</td><td>Whether it works on your build</td></tr><tr><td><a href="https://www.first.org/epss/" target="_blank" rel="noopener noreferrer" class="">EPSS</a></td><td>Probability a published CVE is exploited in the wild in the next 30 days</td><td>Whether your deployment is exposed</td></tr><tr><td><a href="https://www.cisa.gov/known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">CISA KEV</a></td><td>Reliable evidence of active exploitation in the wild, plus a clear remediation action; <a href="https://www.cisa.gov/news-events/directives/bod-26-04-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04</a> says the inclusion criteria "remain unchanged"</td><td>Whether your configuration is exposed</td></tr><tr><td>A demonstrated exploit</td><td>It worked on this build, from this starting position</td><td>Whether attackers are using it, and what else is wrong</td></tr></tbody></table>
<p>The standards already treat exploitation evidence as a variable that matters, but they apply it to the vulnerability out in the world. None has seen your build, configuration or data path. EPSS and KEV are also defined over CVEs, so a flaw in your own code with no CVE gets no signal from either.</p>
<p>My position is that the same logic should run locally. If evidence of exploitation moves a vulnerability up the global list, evidence that it works on your system should move it up yours.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="why-the-evidence-is-worth-the-effort">Why the evidence is worth the effort<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#why-the-evidence-is-worth-the-effort" class="hash-link" aria-label="Direct link to Why the evidence is worth the effort" title="Direct link to Why the evidence is worth the effort" translate="no">​</a></h2>
<p>Two numbers look like they conflict. CISA's FAQ for <a href="https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">BOD 22-01</a>, a 2021 directive since superseded, cites a study of historical vulnerability data dating back to 2019 and says "less than 4% of all known vulnerabilities have been used by attackers in the wild." <a href="https://www.verizon.com/business/resources/executivebriefs/2026-dbir-executive-summary.pdf" target="_blank" rel="noopener noreferrer" class="">Verizon's 2026 DBIR executive summary</a> says exploitation of vulnerabilities is now the most common initial access vector for breaches, at 31% in its reporting dataset, with credential abuse down to 13%.</p>
<p>One number counts vulnerabilities and the other counts breaches. Most known vulnerabilities have never been seen used by attackers, and the ones that are used are now the leading way in.</p>
<p>Capacity looks like the constraint to me. The same summary says only 26% of critical vulnerabilities, defined as those in CISA's KEV catalog, were fully remediated in 2025, down from 38% the previous year. The median time to full resolution went up to 43 days from 32, and in the median case organizations had 50% more critical vulnerabilities to patch than the year before. A finding with nothing behind it competes for what is left and loses.</p>
<p>"It is only theoretical" is also weaker than it sounds. <a href="https://cloud.google.com/blog/topics/threat-intelligence/time-to-exploit-trends-2023" target="_blank" rel="noopener noreferrer" class="">Mandiant</a> put average time to exploit at five days in 2023 with outliers removed, 47 days with them left in, and has "continued not to observe a correlation" between exploit availability and exploitation timing. No public proof of concept is not a safety argument.</p>
<p>Unproven claims have a cost too, and the receiver pays it. curl <a href="https://www.theregister.com/2026/01/21/curl_ends_bug_bounty/" target="_blank" rel="noopener noreferrer" class="">ended its bug bounty</a> at the end of January 2026 to "remove the incentive for people to submit crap and non-well researched reports to us." The Register reported that none of the seven reports from the most recent week described a security vulnerability. A scanner backlog is not a bounty inbox, but the economics match: the claim is cheap and checking it is expensive.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-counts-as-proof">What counts as proof<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#what-counts-as-proof" class="hash-link" aria-label="Direct link to What counts as proof" title="Direct link to What counts as proof" translate="no">​</a></h2>
<p>NIST's <a href="https://csrc.nist.gov/glossary/term/penetration_testing" target="_blank" rel="noopener noreferrer" class="">glossary</a> sets the bar: penetration testing "often involves issuing real attacks on real systems and data," and most tests look for combinations of vulnerabilities that give more access than one alone. Proof is an active attempt, not an inference.</p>
<p>I hold a proof to four questions. On what did it work: the deployed build and configuration, not the main branch or a lab image sharing a version number. From where did it start: the position a real attacker would hold, stated plainly. What happened: an observed effect, and a chain counts. Can someone else rerun it: the owner should be able to check it, and repeat it after the fix.</p>
<p>A version match is not proof. Neither is a public proof of concept that ran in a lab. Those are reasons to run the test. Write "demonstrated" in the ticket and keep "exploited" for evidence of real attackers.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="proving-it-without-causing-the-incident">Proving it without causing the incident<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#proving-it-without-causing-the-incident" class="hash-link" aria-label="Direct link to Proving it without causing the incident" title="Direct link to Proving it without causing the incident" translate="no">​</a></h2>
<p>"Real attacks on real systems and data" is exactly what makes this dangerous. Agree the limits first, in writing: which systems, which window, who is testing, when to stop. If a third party owns part of the path, their permission is a separate question.</p>
<p>Then keep the attempt non-destructive. Prove a read with a canary record you planted, not customer data. Prove execution with a harmless marker. Stop at the first evidence and leave nothing behind. Prefer a pre-production copy of the same build and configuration, and record how it differs from production, because each difference weakens the result.</p>
<p>In the US, TechCrunch reported in May 2022 that the Justice Department's <a href="https://techcrunch.com/2022/05/19/justice-department-good-fatih-hackers-cfaa/" target="_blank" rel="noopener noreferrer" class="">revised CFAA charging policy</a> "directs that good-faith security research should not be charged," for research done "in a manner designed to avoid any harm to individuals or the public." That is charging policy, not a statute, and it does not bind states. Your written authorization is what you rely on, so involve counsel before testing anything you do not clearly own.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-stops-applying">Where this stops applying<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#where-this-stops-applying" class="hash-link" aria-label="Direct link to Where this stops applying" title="Direct link to Where this stops applying" translate="no">​</a></h2>
<p>Proof is asymmetric.</p>

























<table><thead><tr><th>Result</th><th>You can say</th><th>You cannot say</th></tr></thead><tbody><tr><td>It worked</td><td>This build, from this position, is exploitable</td><td>Attackers are using it, or nothing else is wrong</td></tr><tr><td>It failed</td><td>This attempt failed</td><td>The flaw is unreachable</td></tr><tr><td>Not tried</td><td>Nothing</td><td>It is safe</td></tr></tbody></table>
<p>CISA's <a href="https://www.cisa.gov/sites/default/files/2023-01/VEX_Status_Justification_Jun22.pdf" target="_blank" rel="noopener noreferrer" class="">VEX justification guidance</a> makes a related point about code said to be out of the execution path: paraphrasing, do not assume it, because an attacker may be able to divert the path. Treat "could not exploit" as a low confidence signal, never as a reason to close a finding on its own.</p>
<p>Nor should proof become a precondition for fixing things. If a vulnerability is in KEV and you run it, fix it.</p>
<p>It also does not scale. Exploitability depends on environment, chaining and attacker skill, so a proof on one finding generalises poorly to the next, and you will never test the whole backlog. Spend it on disputed findings and expensive fixes.</p>
<p>And it does not replace coverage. You can only test what you know exists, the argument of <a class="" href="https://docs.dwaar.ai/blog/you-cannot-secure-what-you-cannot-list">You Cannot Secure What You Cannot List</a>. Knowing what exists is what Dwaar works on: discovering what you ship from source control, registries and cloud accounts, scanning it, and keeping one findings list with EPSS and KEV signals beside it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-it-on-one-ticket">Try it on one ticket<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#try-it-on-one-ticket" class="hash-link" aria-label="Direct link to Try it on one ticket" title="Direct link to Try it on one ticket" translate="no">​</a></h2>
<p>Pick your oldest open critical where the owner disputes the finding. Check KEV first; if it is listed, stop arguing and fix it. If not, get written approval, make one non-destructive attempt against a pre-production copy of the deployed build, and answer the four questions on the ticket. Mark it with one word: demonstrated, failed or untested. Then see whether that ticket moves faster than the ones around it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/a-finding-is-a-claim-exploitability-is-the-proof#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://www.first.org/cvss/v4-0/specification-document" target="_blank" rel="noopener noreferrer" class="">CVSS v4.0 Specification</a>, FIRST, 2024-06-18</li>
<li class=""><a href="https://certcc.github.io/SSVC/reference/decision_points/exploitation/" target="_blank" rel="noopener noreferrer" class="">SSVC Exploitation decision point</a>, CERT/CC, version ssvc:E:1.1.0, undated page</li>
<li class=""><a href="https://www.first.org/epss/" target="_blank" rel="noopener noreferrer" class="">EPSS</a>, FIRST, undated</li>
<li class=""><a href="https://www.cisa.gov/known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">Known Exploited Vulnerabilities</a>, CISA, undated</li>
<li class=""><a href="https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities" target="_blank" rel="noopener noreferrer" class="">BOD 22-01</a>, CISA, 2021-11-03</li>
<li class=""><a href="https://www.cisa.gov/news-events/directives/bod-26-04-prioritizing-security-updates-based-risk" target="_blank" rel="noopener noreferrer" class="">BOD 26-04: Prioritizing Security Updates Based on Risk</a>, CISA, 2026-06-10</li>
<li class=""><a href="https://www.verizon.com/business/resources/executivebriefs/2026-dbir-executive-summary.pdf" target="_blank" rel="noopener noreferrer" class="">2026 DBIR Executive Summary</a>, Verizon, 2026; the report was <a href="https://www.helpnetsecurity.com/?p=371648" target="_blank" rel="noopener noreferrer" class="">out by 2026-05-25</a>, and this PDF is dated 2026-06-05</li>
<li class=""><a href="https://cloud.google.com/blog/topics/threat-intelligence/time-to-exploit-trends-2023" target="_blank" rel="noopener noreferrer" class="">Time-to-Exploit Trends 2023</a>, Google Cloud / Mandiant, 2024-10-16</li>
<li class=""><a href="https://www.theregister.com/2026/01/21/curl_ends_bug_bounty/" target="_blank" rel="noopener noreferrer" class="">Curl shutters bug bounty program</a>, The Register, 2026-01-21</li>
<li class=""><a href="https://csrc.nist.gov/glossary/term/penetration_testing" target="_blank" rel="noopener noreferrer" class="">Penetration testing glossary entry</a>, NIST, 2008-09</li>
<li class=""><a href="https://techcrunch.com/2022/05/19/justice-department-good-fatih-hackers-cfaa/" target="_blank" rel="noopener noreferrer" class="">DOJ on good-faith hackers and the CFAA</a>, TechCrunch, 2022-05-19</li>
<li class=""><a href="https://www.cisa.gov/sites/default/files/2023-01/VEX_Status_Justification_Jun22.pdf" target="_blank" rel="noopener noreferrer" class="">VEX Status Justification</a>, CISA, 2022-06</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>Red teaming</category>
            <category>Vulnerability management</category>
        </item>
        <item>
            <title><![CDATA[ContextForge CVE-2026-53710 and Chainlit CVE-2026-45018: Allowlists and Denylists Around MCP Keep Failing]]></title>
            <link>https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit</link>
            <guid>https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit</guid>
            <pubDate>Wed, 26 Aug 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[A sandbox escape in ContextForge's python_sandbox_server and an unauthenticated command injection in Chainlit's /mcp endpoint both beat text checks. Fixed in 1.0.2 and 2.12.0.]]></description>
            <content:encoded><![CDATA[<p>Somewhere in your agent stack, a piece of code decides what may run, and it decides by reading text: a binary name, a string of Python. Two advisories published this week, one day apart, show what that check is worth. If you run Chainlit with MCP turned on, or ContextForge's Python sandbox server over HTTP, patch today.</p>








































<table><thead><tr><th>At a glance</th><th>ContextForge <code>python_sandbox_server</code></th><th>Chainlit <code>/mcp</code></th></tr></thead><tbody><tr><td>Identifier</td><td>CVE-2026-53710, <a href="https://github.com/IBM/mcp-context-forge/security/advisories/GHSA-xm98-3vcf-fph7" target="_blank" rel="noopener noreferrer" class="">GHSA-xm98-3vcf-fph7</a></td><td>CVE-2026-45018, <a href="https://github.com/Chainlit/chainlit/security/advisories/GHSA-w3fx-mc44-mf6j" target="_blank" rel="noopener noreferrer" class="">GHSA-w3fx-mc44-mf6j</a></td></tr><tr><td>Affected</td><td>pip <code>mcp-contextforge-gateway</code> <code>&lt;= 1.0.1</code> per the advisory, this sub-project only</td><td>pip <code>chainlit</code> <code>&gt;= 2.4.0rc0, &lt; 2.12.0</code>, only with <code>features.mcp.enabled = true</code></td></tr><tr><td>Fixed in</td><td><code>&gt;= 1.0.2</code> per the advisory; the fix is in the v1.0.2 source, not the PyPI wheel</td><td>2.12.0</td></tr><tr><td>Severity</td><td>Critical in the GitHub advisory, no scorer named: 9.0 HTTP, 8.8 stdio floor</td><td>9.8 Critical in the GitHub advisory, no scorer named; <a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-45018" target="_blank" rel="noopener noreferrer" class="">NVD</a> repeats it, sourced from GitHub</td></tr><tr><td>Exploited as of 2026-08-26</td><td>None reported, PoC in advisory, not in <a href="https://www.cisa.gov/known-exploited-vulnerabilities-catalog" target="_blank" rel="noopener noreferrer" class="">CISA KEV</a></td><td>None reported, not in CISA KEV; CISA SSVC on the NVD record says <code>poc</code>, automatable</td></tr><tr><td>Disclosed</td><td>2026-08-24, fix in the v1.0.2 release since 2026-05-26</td><td>2026-08-25, reported 2026-04-08 by external researchers</td></tr></tbody></table>
<p>Keep the scope straight: IBM's advisory says "The core ContextForge gateway and proxy components are not directly affected," and Chainlit's MCP support has been off by default since 2.7.0.</p>
<p>The sources disagree twice. ContextForge's advisory says 9.0, but its listed vector (<code>CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H</code>) computes to 10.0, and NVD has no entry yet. The Critical rating assumes HTTP reachability, so your transport matters more than the decimal. Chainlit's sidebar says <code>&lt;= 2.11.1</code> with no patch, while its body says <code>&lt; 2.12.0</code>, patched in 2.12.0. <a href="https://pypi.org/pypi/chainlit/2.12.0/json" target="_blank" rel="noopener noreferrer" class="">PyPI</a> has 2.12.0 from 2026-08-25, so go with the body.</p>
<p>One more trap on the ContextForge side. The advisory's affected package is pip <code>mcp-contextforge-gateway</code>, but the file it names as the component is not in that package. <a href="https://github.com/IBM/mcp-context-forge/blob/v1.0.2/pyproject.toml" target="_blank" rel="noopener noreferrer" class="">The gateway's packaging config</a> keeps only packages starting with <code>mcpgateway</code>, and <a href="https://pypi.org/pypi/mcp-contextforge-gateway/1.0.2/json" target="_blank" rel="noopener noreferrer" class="">the 1.0.2 wheel and sdist on PyPI</a> have no sandbox path. The sandbox has <a href="https://github.com/IBM/mcp-context-forge/blob/v1.0.2/mcp-servers/python/python_sandbox_server/pyproject.toml" target="_blank" rel="noopener noreferrer" class="">its own <code>pyproject.toml</code></a> as <code>python-sandbox-server</code> 2.0.0, and <a href="https://github.com/IBM/mcp-context-forge/blob/v1.0.1/mcp-servers/python/python_sandbox_server/pyproject.toml" target="_blank" rel="noopener noreferrer" class="">v1.0.1 says the same</a>, so its version number cannot tell you whether you have the fix.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="chainlit-an-allowlist-that-read-one-word">Chainlit: an allowlist that read one word<a href="https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit#chainlit-an-allowlist-that-read-one-word" class="hash-link" aria-label="Direct link to Chainlit: an allowlist that read one word" title="Direct link to Chainlit: an allowlist that read one word" translate="no">​</a></h2>
<p><code>POST /mcp</code> accepted a client-supplied <code>fullCommand</code> string. Per <a href="https://github.com/Chainlit/chainlit/security/advisories/GHSA-w3fx-mc44-mf6j" target="_blank" rel="noopener noreferrer" class="">the advisory</a>, <code>validate_mcp_command()</code> split it with <code>shlex.split()</code>, checked only the executable name, and passed the arguments straight to <code>StdioServerParameters</code>. <code>npx</code> looks safe to allow, and it supports <code>-c</code>, which runs a shell command. The check passes and the attacker chooses the rest.</p>
<p>Three details make it worse:</p>
<ul>
<li class=""><code>allowed_executables</code> defaults to <code>None</code>, and the code treats <code>None</code> as "allow everything."</li>
<li class="">The route is always registered. Only the runtime flag and any configured auth guard it, and without an auth callback a client needs just a session id, with no login.</li>
<li class=""><a href="https://github.com/Chainlit/chainlit/blob/2.12.0/docs/security-advisory-2026-mcp.md" target="_blank" rel="noopener noreferrer" class="">Chainlit's write-up</a> says: "The server spawned the subprocess before the MCP handshake completed, so the shell command executed even though the MCP connection subsequently failed." A failed connection in your logs proves nothing.</li>
</ul>
<p>It also gives the verdict: "No argument-level filtering can safely sandbox this because the attacker controls the full argument list." So 2.12.0 removes the input. A client used to send a command line. Now it sends a session id and a name, and stdio servers are declared server-side under <code>[[features.mcp.servers]]</code>.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="contextforge-a-denylist-that-read-the-source">ContextForge: a denylist that read the source<a href="https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit#contextforge-a-denylist-that-read-the-source" class="hash-link" aria-label="Direct link to ContextForge: a denylist that read the source" title="Direct link to ContextForge: a denylist that read the source" translate="no">​</a></h2>
<p>The <code>execute_code</code> tool runs user Python under RestrictedPython. <a href="https://github.com/IBM/mcp-context-forge/security/advisories/GHSA-xm98-3vcf-fph7" target="_blank" rel="noopener noreferrer" class="">The advisory</a> lists three weaknesses: raw <code>getattr</code> in <code>safe_builtins</code>, <code>validate_code</code> looking for literal dunder strings, and <code>execute_code</code> being exposable over HTTP/SSE with no authentication.</p>
<p><a href="https://github.com/IBM/mcp-context-forge/commit/d6c8c614414246a321784cf45c6b3aa99bda9563" target="_blank" rel="noopener noreferrer" class="">The fix commit</a> counts four, adding that the <code>_getattr_</code> guard was missing from globals and that the builtins merge order let raw <code>getattr</code> overwrite RestrictedPython's entries. On the validator it is blunt: "trivially bypassed by <code>'__cla' + 'ss__'</code>". The advisory's PoC builds dunder names at runtime and walks the class hierarchy to <code>subprocess.Popen</code>, while validation reports <code>'valid': True</code>.</p>
<p>Anyone who can reach the endpoint gets OS commands with the server process's privileges. The advisory warns of filesystem compromise or pivoting where a container has host mounts or internal reachability.</p>
<p>Now the dates. <a href="https://github.com/IBM/mcp-context-forge/releases/tag/v1.0.2" target="_blank" rel="noopener noreferrer" class="">The v1.0.2 release</a> was published on 2026-05-26. Its notes mention other hardening but not the sandbox, even though the fix commit in that release, dated 2026-05-25, already names the GHSA id and describes the escape. The advisory came about three months later, and until then anything deciding upgrades from advisories had nothing to match.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-check-and-do-now">What to check and do now<a href="https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit#what-to-check-and-do-now" class="hash-link" aria-label="Direct link to What to check and do now" title="Direct link to What to check and do now" translate="no">​</a></h2>
<ol>
<li class=""><strong>ContextForge.</strong> Find every deployment of <code>python_sandbox_server</code>, and treat <code>mcp-contextforge-gateway</code> <code>&lt;= 1.0.1</code> as the advisory's version marker for it. The sandbox server is not in the PyPI wheel, so upgrading the gateway with pip does not patch it. Redeploy it from v1.0.2 or later source. The code lives at <code>mcp-servers/python/python_sandbox_server</code>, so I would check source checkouts and images too.</li>
<li class=""><strong>Set the token first.</strong> Over HTTP, fixed versions require <code>SANDBOX_API_TOKEN</code>, at least 32 characters, sent as a bearer token. The server exits if it is unset.</li>
<li class=""><strong>Chainlit.</strong> Find <code>chainlit</code> <code>&gt;= 2.4.0rc0, &lt; 2.12.0</code> and read <code>.chainlit/config.toml</code>. The off default only dates from 2.7.0, so do not assume it.</li>
<li class=""><strong>Upgrade to 2.12.0 and migrate.</strong> Legacy <code>[features.mcp.sse]</code>, <code>[features.mcp.stdio]</code> and <code>[features.mcp.streamable-http]</code> sections, or <code>allowed_executables</code>, stop the app starting once MCP is enabled. This release also fixes CVE-2026-45019, an SSRF scored 7.2.</li>
<li class=""><strong>If you cannot upgrade today,</strong> set <code>features.mcp.enabled = false</code>, which covers both Chainlit CVEs. An auth callback narrows who can reach <code>/mcp</code> but does not fix the injection. Block egress to internal networks and cloud metadata.</li>
<li class=""><strong>Add auth after upgrading too.</strong> Per the advisory, 2.12.0 with no auth still lets anonymous clients spawn your configured stdio servers by name, with no resource limits or per-client session cap.</li>
<li class=""><strong>Look back.</strong> Neither advisory gives indicators of compromise. My inference, not sourced guidance: search for <code>POST /mcp</code> bodies containing <code>fullCommand</code>, and for unexpected <code>npx</code> or <code>sh</code> children of either server process.</li>
</ol>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-enforce-instead">What to enforce instead<a href="https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit#what-to-enforce-instead" class="hash-link" aria-label="Direct link to What to enforce instead" title="Direct link to What to enforce instead" translate="no">​</a></h2>
<p><a class="" href="https://docs.dwaar.ai/blog/mcp-stdio-command-execution-by-design">Stdio makes a server definition a process launch</a>, so whoever influences that definition chooses what your host runs. Chainlit policed that by reading command text, and ContextForge's validator did the same to Python source. ContextForge's fix adds a dunder allowlist of its own, enforced on actual attribute access. Lists are fine. What fails is a list applied to a string the attacker wrote.</p>
<p>What I would hold to in any agent stack:</p>
<ul>
<li class=""><strong>Callers pick a name, never a command.</strong> Operators declare commands server-side, and no request body carries a command or its arguments.</li>
<li class=""><strong>Authenticate before you spawn, and fail closed.</strong> Chainlit's <code>None</code> meant allow everything, and ContextForge kept a fallback <code>eval</code>/<code>exec</code> path for when RestrictedPython was unavailable. Exiting on a missing token is the behaviour to copy.</li>
<li class=""><strong>Assume the language sandbox falls.</strong> AST checks and guarded attribute access beat substring matching, and I still would not make them the boundary. <a class="" href="https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem">vm2 last week</a> is the same story in another language. Run code execution tools with no host mounts, no credentials and no route to internal networks.</li>
<li class=""><strong>Treat config and transport as part of the vulnerability.</strong> A version match flags a ContextForge gateway install that does not contain the sandbox server and a Chainlit app with MCP off, misses a sandbox server whose own version did not change with the fix, and told you nothing during ContextForge's three month gap.</li>
</ul>
<p>Today, search your repositories and images for <code>chainlit</code>, <code>mcp-contextforge-gateway</code>, <code>python_sandbox_server</code> and <code>python-sandbox-server</code>. For each hit, name the authentication check that runs before any process starts. If you cannot name one, that is your next fix.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/mcp-tool-execution-rce-in-contextforge-and-chainlit#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">IBM (GitHub Security Advisory), <a href="https://github.com/IBM/mcp-context-forge/security/advisories/GHSA-xm98-3vcf-fph7" target="_blank" rel="noopener noreferrer" class="">GHSA-xm98-3vcf-fph7</a>, 2026-08-24</li>
<li class="">IBM (GitHub commit), <a href="https://github.com/IBM/mcp-context-forge/commit/d6c8c614414246a321784cf45c6b3aa99bda9563" target="_blank" rel="noopener noreferrer" class="">d6c8c61</a>, 2026-05-25</li>
<li class="">IBM (GitHub release), <a href="https://github.com/IBM/mcp-context-forge/releases/tag/v1.0.2" target="_blank" rel="noopener noreferrer" class="">v1.0.2</a>, 2026-05-26</li>
<li class="">IBM (GitHub repository), <a href="https://github.com/IBM/mcp-context-forge/blob/v1.0.2/pyproject.toml" target="_blank" rel="noopener noreferrer" class="">gateway pyproject.toml at v1.0.2</a>, 2026-05-26</li>
<li class="">IBM (GitHub repository), <a href="https://github.com/IBM/mcp-context-forge/blob/v1.0.2/mcp-servers/python/python_sandbox_server/pyproject.toml" target="_blank" rel="noopener noreferrer" class="">python_sandbox_server pyproject.toml at v1.0.2</a>, 2026-05-26</li>
<li class="">IBM (GitHub repository), <a href="https://github.com/IBM/mcp-context-forge/blob/v1.0.1/mcp-servers/python/python_sandbox_server/pyproject.toml" target="_blank" rel="noopener noreferrer" class="">python_sandbox_server pyproject.toml at v1.0.1</a>, 2026-05-13</li>
<li class="">PyPI, <a href="https://pypi.org/pypi/mcp-contextforge-gateway/1.0.2/json" target="_blank" rel="noopener noreferrer" class="">mcp-contextforge-gateway 1.0.2 metadata</a>, uploaded 2026-05-26</li>
<li class="">Chainlit (GitHub Security Advisory), <a href="https://github.com/Chainlit/chainlit/security/advisories/GHSA-w3fx-mc44-mf6j" target="_blank" rel="noopener noreferrer" class="">GHSA-w3fx-mc44-mf6j</a>, 2026-08-25</li>
<li class="">Chainlit (GitHub repository), <a href="https://github.com/Chainlit/chainlit/blob/2.12.0/docs/security-advisory-2026-mcp.md" target="_blank" rel="noopener noreferrer" class="">Security Advisory: MCP Endpoint Vulnerabilities</a>, 2026-08-25</li>
<li class="">PyPI, <a href="https://pypi.org/pypi/chainlit/2.12.0/json" target="_blank" rel="noopener noreferrer" class="">chainlit 2.12.0 metadata</a>, uploaded 2026-08-25</li>
<li class="">NVD, <a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-45018" target="_blank" rel="noopener noreferrer" class="">CVE-2026-45018</a>, 2026-08-25</li>
<li class="">CISA, <a href="https://www.cisa.gov/known-exploited-vulnerabilities-catalog" target="_blank" rel="noopener noreferrer" class="">Known Exploited Vulnerabilities Catalog</a>, living catalog</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>AI agents</category>
            <category>MCP</category>
        </item>
        <item>
            <title><![CDATA[The Case for a Dependency Cooldown]]></title>
            <link>https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown</link>
            <guid>https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown</guid>
            <pubDate>Thu, 20 Aug 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[In one analysis of ten supply chain attacks, eight were removed in under a week. Why waiting before adopting a new version works, where it fails, and how to exempt security fixes.]]></description>
            <content:encoded><![CDATA[<p>An update bot with no cooldown will open a pull request for a release published minutes ago, and your CI will install it with whatever credentials that job holds. A hijacked maintainer account needs exactly that speed.</p>
<p>The <a href="https://github.com/axios/axios/issues/10636" target="_blank" rel="noopener noreferrer" class="">axios postmortem</a> puts the malicious 1.14.1 on npm at 00:21 UTC on March 31, 2026, with both it and 0.30.4 gone by 03:15 UTC (<a class="" href="https://docs.dwaar.ai/blog/axios-npm-hijack-maintainer-social-engineering">our earlier post</a> covers how). About three hours. A team refusing any version younger than a day would never have resolved either one.</p>
<p>My position: do not adopt a new version of a third-party dependency until it has been public for a while, and make that configuration, not discipline. William Woodruff <a href="https://blog.yossarian.net/2025/11/21/We-should-all-be-using-dependency-cooldowns" target="_blank" rel="noopener noreferrer" class="">argued this in November 2025</a>, and GitHub <a href="https://github.blog/security/supply-chain-security/the-case-for-a-cooldown-why-dependabot-now-waits-before-issuing-version-updates/" target="_blank" rel="noopener noreferrer" class="">made much the same case in July</a>. The name is a dependency cooldown.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="why-the-obvious-answers-fall-short">Why the obvious answers fall short<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#why-the-obvious-answers-fall-short" class="hash-link" aria-label="Direct link to Why the obvious answers fall short" title="Direct link to Why the obvious answers fall short" translate="no">​</a></h2>
<p>The obvious advice is to update fast. That advice was built for vulnerable code, where the newest version is the safest one available. A malicious release inverts it, because the newest version is the dangerous one. <a class="" href="https://docs.dwaar.ai/blog/a-vulnerable-package-and-a-malicious-package-are-different-problems">These are different problems</a>.</p>
<p>The other obvious answer is account security. CISA's <a href="https://www.cisa.gov/news-events/alerts/2025/09/23/widespread-supply-chain-compromise-impacting-npm-ecosystem" target="_blank" rel="noopener noreferrer" class="">alert on the Shai-Hulud npm worm</a> advises phishing-resistant MFA, and it should. But you do not run those accounts, and the axios account had 2FA enabled when a targeted social engineering campaign reached its lead maintainer.</p>
<p>That same alert says to pin to known safe releases produced before September 16, 2025. I read that as an age rule applied after the fact. A cooldown applies it up front.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-the-wait-buys">What the wait buys<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#what-the-wait-buys" class="hash-link" aria-label="Direct link to What the wait buys" title="Direct link to What the wait buys" translate="no">​</a></h2>
<p>A cooldown detects nothing. It moves your adoption date. It works because the version stays public during the wait, which, in GitHub's words, "gives maintainers, security researchers, and automated scanners time to spot a malicious version."</p>
<p>The most useful public data is <a href="https://blog.yossarian.net/2025/11/21/We-should-all-be-using-dependency-cooldowns" target="_blank" rel="noopener noreferrer" class="">Woodruff's table</a> of ten attacks, measured by window of opportunity: the attacker's public release to removal upstream. That is not time to detection, but it is how long the bad version was available.</p>






























<table><thead><tr><th>Cooldown</th><th>Windows it outlasts</th><th>Still gets through</th></tr></thead><tbody><tr><td>1 day</td><td>Seven of ten: Ultralytics (12 hours, then 1 hour), chalk (under 12 hours), Nx (4 hours), rspack (1 hour), num2words (under 12 hours), web3.js (5 hours)</td><td>tj-actions, Kong Ingress Controller, xz-utils</td></tr><tr><td>3 days</td><td>The same seven, with tj-actions (about 3 days) too close to call</td><td>Kong, xz-utils</td></tr><tr><td>7 days</td><td>Eight of ten, adding tj-actions</td><td>Kong (about 10 days), xz-utils</td></tr><tr><td>14 days</td><td>Nine of ten, adding Kong</td><td>xz-utils (about 5 weeks)</td></tr></tbody></table>
<p>That is ten incidents on a practitioner's blog, not a rate across all malicious packages. <a href="https://github.blog/security/supply-chain-security/the-case-for-a-cooldown-why-dependabot-now-waits-before-issuing-version-updates/" target="_blank" rel="noopener noreferrer" class="">GitHub</a> cites one developer's survey of 21 incidents, which its author says is not a controlled study, in which the malicious axios, Solana web3.js, ua-parser-js and Ledger Connect Kit versions were each pulled within hours. GitHub makes Dependabot and has a stake in that conclusion. Sources disagree on details: chalk is under 12 hours for Woodruff, while GitHub says the September 2025 versions were live roughly two hours. Both are under a day. tj-actions is about 3 days for Woodruff, while <a href="https://nvd.nist.gov/vuln/detail/CVE-2025-30066" target="_blank" rel="noopener noreferrer" class="">NVD</a> lists the tags as affected on March 14 and 15, 2025.</p>
<p>The academic number I can point to looks very different. <a href="https://arxiv.org/abs/2005.09535" target="_blank" rel="noopener noreferrer" class="">Ohm et al.</a> collected 174 malicious npm, PyPI and RubyGems packages dating from November 2015 to November 2019, and report that a package was available for 209 days on average before being publicly reported, with a median of 67. But 61% of those packages were typosquats, not hijacked releases of popular projects, the clock stops at the advisory rather than at removal, and the data ends in 2019. I read it as a warning not to stretch this argument to name-squatting malware, which can outlast any wait you would accept. For hijacked releases of popular packages, I have not found an independent, population-level measurement.</p>
<p>My number is seven days. Three is the floor: it has been <a href="https://github.blog/changelog/2026-07-14-dependabot-version-updates-introduce-default-package-cooldown/" target="_blank" rel="noopener noreferrer" class="">Dependabot's default since July 14</a>. Seven clears tj-actions on any of these dates, which three may not. Fourteen adds one row for another week behind, so I would save it for major versions.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-settings-already-exist">The settings already exist<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#the-settings-already-exist" class="hash-link" aria-label="Direct link to The settings already exist" title="Direct link to The settings already exist" translate="no">​</a></h2>
<p>This is what each tool's documentation describes. Check the page for your version before copying anything.</p>






























<table><thead><tr><th>Tool</th><th>Setting</th><th>Escape hatch or edge</th></tr></thead><tbody><tr><td><a href="https://docs.github.com/en/code-security/dependabot/working-with-dependabot/dependabot-options-reference" target="_blank" rel="noopener noreferrer" class="">Dependabot</a></td><td><code>cooldown</code>: <code>default-days</code>, <code>semver-major-days</code></td><td><code>exclude</code> list, which beats <code>include</code></td></tr><tr><td><a href="https://pnpm.io/settings/dependency-resolution" target="_blank" rel="noopener noreferrer" class="">pnpm</a></td><td><code>minimumReleaseAge</code>, in minutes</td><td><code>minimumReleaseAgeExclude</code> lists package names</td></tr><tr><td><a href="https://docs.npmjs.com/cli/v11/using-npm/config" target="_blank" rel="noopener noreferrer" class="">npm</a></td><td><code>min-release-age</code>, in days, <a href="https://github.com/npm/cli/releases/tag/v11.10.0" target="_blank" rel="noopener noreferrer" class="">since CLI 11.10.0</a></td><td><code>min-release-age-exclude</code> lists package names or globs (<a href="https://github.com/npm/cli/releases/tag/v11.17.0" target="_blank" rel="noopener noreferrer" class="">since 11.17.0</a>); errors if no eligible version exists</td></tr><tr><td><a href="https://docs.astral.sh/uv/reference/settings/#exclude-newer" target="_blank" rel="noopener noreferrer" class="">uv</a></td><td><code>exclude-newer</code>, a date or a duration such as 1 week</td><td><code>exclude-newer-package</code> set to false</td></tr></tbody></table>
<p>Bot settings govern what gets proposed, and installer settings govern what gets installed. A Dependabot setting says nothing about a developer's manual install, so use both where you can.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-exception-for-security-patches">The exception for security patches<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#the-exception-for-security-patches" class="hash-link" aria-label="Direct link to The exception for security patches" title="Direct link to The exception for security patches" translate="no">​</a></h2>
<p>The serious objection is that seven days is a long time to sit on a fix for a public flaw. Correct, and the answer is a separate path, not a shorter cooldown.</p>
<p>Dependabot already draws this line. Its reference says "The cooldown option is only available for version updates, not security updates."</p>
<p>Two gaps remain. A fix for a vulnerability with no published advisory yet is an ordinary version update, so it waits. And at the installer level the exception is yours to build: pnpm, npm and uv all give you per-package exemptions you add by hand. npm errors when nothing is eligible, and <a href="https://github.com/npm/cli/releases/tag/v11.18.0" target="_blank" rel="noopener noreferrer" class="">since 11.18.0</a>, if the window blocks an <code>npm audit fix</code> patch, it keeps the vulnerable version, warns and exits non-zero. Find out what your emergency path does before you need it.</p>
<p>The rule I would write down: bypass only when an advisory from a database you trust names the version you currently lock. Do not bypass because the new version's release notes say "security fix", since that is what an attacker would write. Remove the exemption once the upgrade lands.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-stops-working">Where this stops working<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#where-this-stops-working" class="hash-link" aria-label="Direct link to Where this stops working" title="Direct link to Where this stops working" translate="no">​</a></h2>
<p><a href="https://www.cisa.gov/news-events/alerts/2024/03/29/reported-supply-chain-compromise-affecting-xz-utils-data-compression-library-cve-2024-3094" target="_blank" rel="noopener noreferrer" class="">CISA</a> names xz-utils 5.6.0 and 5.6.1 as carrying malicious code, and Woodruff measures about five weeks from the first backdoored release. GitHub says a cooldown "does little against attacks that play a longer game", such as dormant backdoors, maintainer sabotage or a compromised build system.</p>
<p>Against a worm like Shai-Hulud, it helps only if each release is removed inside the wait. Nor does it touch a bad version already in your lockfile.</p>
<p>Two questions stay open, and these are my inferences, not measurements. If everyone waits, who finds the bad version? Woodruff's answer is that public indices get scanned regardless, which is plausible and unproven at scale. And once cooldowns are common, patient attackers have a reason to hold a payload past the window.</p>
<p>A cooldown still takes away the cheap, fast, hijacked-account attack and leaves the rest for your other controls.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-it-on-one-repository">Try it on one repository<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#try-it-on-one-repository" class="hash-link" aria-label="Direct link to Try it on one repository" title="Direct link to Try it on one repository" translate="no">​</a></h2>
<p>Pick the repository with the most automated updates and go through the dependency bumps merged in the last month. Compare each version's publish date with your merge date, and count how many went in inside seven days. Then walk your most recent security update through a seven-day rule: which path would it take, and who would approve the bypass? If nobody can answer, write that rule first and set the cooldown second.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/the-case-for-a-dependency-cooldown#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">William Woodruff, ENOSUCHBLOG, <a href="https://blog.yossarian.net/2025/11/21/We-should-all-be-using-dependency-cooldowns" target="_blank" rel="noopener noreferrer" class="">We should all be using dependency cooldowns</a>, November 21, 2025</li>
<li class="">GitHub Blog, <a href="https://github.blog/security/supply-chain-security/the-case-for-a-cooldown-why-dependabot-now-waits-before-issuing-version-updates/" target="_blank" rel="noopener noreferrer" class="">The case for a cooldown</a>, July 23, 2026</li>
<li class="">GitHub Changelog, <a href="https://github.blog/changelog/2026-07-14-dependabot-version-updates-introduce-default-package-cooldown/" target="_blank" rel="noopener noreferrer" class="">Default package cooldown</a>, July 14, 2026</li>
<li class="">GitHub Docs, <a href="https://docs.github.com/en/code-security/dependabot/working-with-dependabot/dependabot-options-reference" target="_blank" rel="noopener noreferrer" class="">Dependabot options reference</a>, as of August 2026</li>
<li class="">pnpm, <a href="https://pnpm.io/settings/dependency-resolution" target="_blank" rel="noopener noreferrer" class="">Dependency resolution settings</a>, as of August 2026</li>
<li class="">npm, <a href="https://docs.npmjs.com/cli/v11/using-npm/config" target="_blank" rel="noopener noreferrer" class="">npm config</a>, CLI 11.19.0, July 29, 2026</li>
<li class="">npm/cli, <a href="https://github.com/npm/cli/releases/tag/v11.10.0" target="_blank" rel="noopener noreferrer" class="">v11.10.0 release</a>, February 11, 2026</li>
<li class="">npm/cli, <a href="https://github.com/npm/cli/releases/tag/v11.17.0" target="_blank" rel="noopener noreferrer" class="">v11.17.0 release</a>, June 11, 2026</li>
<li class="">npm/cli, <a href="https://github.com/npm/cli/releases/tag/v11.18.0" target="_blank" rel="noopener noreferrer" class="">v11.18.0 release</a>, June 29, 2026</li>
<li class="">Astral, <a href="https://docs.astral.sh/uv/reference/settings/#exclude-newer" target="_blank" rel="noopener noreferrer" class="">uv settings</a>, as of August 2026</li>
<li class="">axios project, <a href="https://github.com/axios/axios/issues/10636" target="_blank" rel="noopener noreferrer" class="">Compromise postmortem</a>, opened April 2, 2026</li>
<li class="">CISA, <a href="https://www.cisa.gov/news-events/alerts/2025/09/23/widespread-supply-chain-compromise-impacting-npm-ecosystem" target="_blank" rel="noopener noreferrer" class="">npm ecosystem compromise alert</a>, September 23, 2025</li>
<li class="">CISA, <a href="https://www.cisa.gov/news-events/alerts/2024/03/29/reported-supply-chain-compromise-affecting-xz-utils-data-compression-library-cve-2024-3094" target="_blank" rel="noopener noreferrer" class="">XZ Utils alert</a>, March 29, 2024</li>
<li class="">NVD, <a href="https://nvd.nist.gov/vuln/detail/CVE-2025-30066" target="_blank" rel="noopener noreferrer" class="">CVE-2025-30066</a>, published March 15, 2025</li>
<li class="">Marc Ohm, Henrik Plate, Arnold Sykosch and Michael Meier, <a href="https://arxiv.org/abs/2005.09535" target="_blank" rel="noopener noreferrer" class="">Backstabber's Knife Collection: A Review of Open Source Software Supply Chain Attacks</a>, May 19, 2020</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>Supply chain</category>
            <category>Open source</category>
        </item>
        <item>
            <title><![CDATA[vm2 Again: Sandbox Escapes and the Incomplete Fix Problem]]></title>
            <link>https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem</link>
            <guid>https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem</guid>
            <pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[vm2 3.11.6 closes two critical sandbox escapes, CVE-2026-47686 and CVE-2026-47698. The second bypasses a fix that shipped in May, so upgrading then did not close it.]]></description>
            <content:encoded><![CDATA[<p>If any of your services runs code it did not write inside vm2, upgrade to 3.11.6 today. The release closes two critical escapes from the sandbox to the host, and one of them gets past a fix the project shipped in May. If you upgraded for that fix, you did the right thing and were still exposed.</p>








































<table><thead><tr><th>At a glance</th><th>CVE-2026-47686</th><th>CVE-2026-47698</th></tr></thead><tbody><tr><td>Advisory</td><td><a href="https://github.com/patriksimek/vm2/security/advisories/GHSA-m283-3h24-438v" target="_blank" rel="noopener noreferrer" class="">GHSA-m283-3h24-438v</a></td><td><a href="https://github.com/patriksimek/vm2/security/advisories/GHSA-cfcw-xp6x-25gj" target="_blank" rel="noopener noreferrer" class="">GHSA-cfcw-xp6x-25gj</a></td></tr><tr><td>Affected</td><td>npm <code>vm2</code> before 3.11.6</td><td>npm <code>vm2</code> before 3.11.6</td></tr><tr><td>Fixed in</td><td>3.11.6</td><td>3.11.6</td></tr><tr><td>Severity</td><td>CVSS 3.1 9.9 Critical, scored by GitHub</td><td>CVSS 3.1 9.8 Critical, scored by GitHub</td></tr><tr><td>Exploited as of 2026-08-18</td><td>No report in the advisory or NVD, PoC in advisory</td><td>No report in the advisory or NVD, PoC in advisory</td></tr><tr><td>Disclosed</td><td>Advisory 2026-08-14, NVD 2026-08-17</td><td>Advisory 2026-08-14, NVD 2026-08-17</td></tr></tbody></table>
<p><a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-47686" target="_blank" rel="noopener noreferrer" class="">NVD</a> lists both scores as Secondary, supplied by GitHub, with no score of its own. The range needs a note too: the text of the 47686 advisory says <code>&lt;= 3.11.3</code>, its package metadata says <code>&lt;= 3.11.5</code>, and NVD says before 3.11.6. Since 3.11.5 listed no security fixes, treat everything before 3.11.6 as affected. <a href="https://github.com/patriksimek/vm2/releases/tag/3.11.6" target="_blank" rel="noopener noreferrer" class="">3.11.6 closes five advisories</a> in all; the others cover <code>os</code> and <code>dns</code> builtin exposure and two <code>bufferAllocLimit</code> denial of service issues.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="two-bugs-with-one-shape">Two bugs with one shape<a href="https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem#two-bugs-with-one-shape" class="hash-link" aria-label="Direct link to Two bugs with one shape" title="Direct link to Two bugs with one shape" translate="no">​</a></h2>
<p>vm2 passes host objects into the sandbox through proxies. The sandbox holds only while raw host objects never reach sandbox code.</p>
<p><strong>CVE-2026-47686 starts with a missed property.</strong> <code>handleException()</code> in <code>lib/setup-sandbox.js</code> walks an error's prototype chain looking for <code>SuppressedError</code> and <code>AggregateError</code>, and recursively sanitises their sub-errors. Any other error is returned without a look at <code>.cause</code>. So if the embedder exposes a host function that throws an error whose cause is a host object such as <code>process</code>, sandbox code catches it, reads <code>e.cause</code>, and can run host commands. The <a href="https://github.com/patriksimek/vm2/security/advisories/GHSA-m283-3h24-438v" target="_blank" rel="noopener noreferrer" class="">advisory</a> adds that <code>docs/ATTACKS.md</code> claimed <code>Error.cause</code> was sanitised.</p>
<p>The <a href="https://github.com/patriksimek/vm2/releases/tag/3.11.6" target="_blank" rel="noopener noreferrer" class="">3.11.6 release notes</a> go further: the <code>SuppressedError</code> and <code>AggregateError</code> slots, arbitrary own properties such as <code>err.detail = process</code>, and the carrier's prototype chain also leaked. The advisory and NVD describe those two slots as sanitised. The <a href="https://github.com/patriksimek/vm2/commit/7e3faaf550f4ab975bf4cdde183fcec49b056d8e" target="_blank" rel="noopener noreferrer" class="">fix commit</a> explains the gap: the read-then-assign sanitising them "silently no-ops against a getter-only accessor".</p>
<p>So search wider than <code>cause</code>. You are looking for any host function you expose that throws an error carrying a host reference. <code>cause</code> is one example:</p>
<div class="language-js codeBlockContainer_Ckt0 theme-code-block" style="--prism-color:var(--ink);--prism-background-color:var(--surface-2)"><div class="codeBlockContent_QJqH"><pre tabindex="0" class="prism-code language-js codeBlock_bY9V thin-scrollbar" style="color:var(--ink);background-color:var(--surface-2)"><code class="codeBlockLines_e6Vv"><div class="token-line" style="color:var(--ink)"><span class="token comment">// Illustration only, not from the advisory.</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token keyword">const</span><span class="token plain"> sandbox </span><span class="token operator">=</span><span class="token plain"> </span><span class="token punctuation">{</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token function">lookup</span><span class="token punctuation">(</span><span class="token parameter">id</span><span class="token punctuation">)</span><span class="token plain"> </span><span class="token punctuation">{</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">    </span><span class="token comment">// The cause is a live host object that sandbox code can catch.</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">    </span><span class="token keyword control-flow">throw</span><span class="token plain"> </span><span class="token keyword">new</span><span class="token plain"> </span><span class="token class-name">Error</span><span class="token punctuation">(</span><span class="token string">"lookup failed"</span><span class="token punctuation">,</span><span class="token plain"> </span><span class="token punctuation">{</span><span class="token plain"> </span><span class="token literal-property property">cause</span><span class="token operator">:</span><span class="token plain"> process </span><span class="token punctuation">}</span><span class="token punctuation">)</span><span class="token punctuation">;</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain">  </span><span class="token punctuation">}</span><span class="token punctuation">,</span><span class="token plain"></span><br></div><div class="token-line" style="color:var(--ink)"><span class="token plain"></span><span class="token punctuation">}</span><span class="token punctuation">;</span><br></div></code></pre></div></div>
<p>The reporter argues this is common because <code>{ cause }</code> chaining is standard in database adapters and HTTP clients. That is an argument, not a measurement.</p>
<p><strong>CVE-2026-47698 is a missed composition.</strong> Host prototype mutators such as <code>__proto__</code> must never reach the sandbox raw, because calling them on a host intrinsic severs its prototype chain and exposes the host <code>Function</code> constructor. <a href="https://github.com/advisories/GHSA-v6mx-mf47-r5wg" target="_blank" rel="noopener noreferrer" class="">The fix for GHSA-v6mx-mf47-r5wg</a>, CVE-2026-47131 with a CVSS of 10.0, shipped in <a href="https://github.com/patriksimek/vm2/releases/tag/v3.11.4" target="_blank" rel="noopener noreferrer" class="">3.11.4 on 18 May</a>. It unwrapped one layer of <code>call</code>, <code>apply</code>, <code>bind</code> or <code>Reflect</code> indirection. <a href="https://github.com/patriksimek/vm2/security/advisories/GHSA-cfcw-xp6x-25gj" target="_blank" rel="noopener noreferrer" class="">The new advisory</a> says plainly that the fix "is insufficient": stack the indirection through <code>Function.prototype.call</code> and the check misses it. Its PoC reaches the mutators on a host <code>Buffer</code>, gets to the host <code>Function</code> constructor and <code>process</code>, and runs a shell command.</p>
<p>The advisory, the <a href="https://github.com/patriksimek/vm2/commit/a85acb61f81402c6eabf32760aa11272af6d0f9e" target="_blank" rel="noopener noreferrer" class="">fix commit</a> and the <a href="https://github.com/patriksimek/vm2/releases/tag/3.11.6" target="_blank" rel="noopener noreferrer" class="">3.11.6 release notes</a> all name GHSA-v6mx-mf47-r5wg as the fix that was bypassed.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-score-points-the-wrong-way">The score points the wrong way<a href="https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem#the-score-points-the-wrong-way" class="hash-link" aria-label="Direct link to The score points the wrong way" title="Direct link to The score points the wrong way" translate="no">​</a></h2>
<p>The 9.9 is the bug with the precondition: it needs a host function you chose to expose. The 9.8 needs only attacker code running inside the sandbox. CISA's SSVC entries on the NVD records, timestamped 18 August, mark <a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-47698" target="_blank" rel="noopener noreferrer" class="">47698</a> as <code>poc</code> and automatable, and 47686 as <code>none</code> and not automatable, although the 47686 advisory attaches a PoC file.</p>
<p>Triage on preconditions, not on the <a class="" href="https://docs.dwaar.ai/blog/cvss-is-not-risk-how-to-use-epss-and-kev">score</a>. The question for any vm2 finding is who writes the code that runs in that sandbox.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="patched-is-not-fixed">Patched is not fixed<a href="https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem#patched-is-not-fixed" class="hash-link" aria-label="Direct link to Patched is not fixed" title="Direct link to Patched is not fixed" translate="no">​</a></h2>
<p>Both bugs are the same failure: a list of dangerous things that was almost complete. A language-level sandbox shares a process with its host, so every object, error and intrinsic is a possible crossing. The defender has to enumerate all of them, and the attacker needs one.</p>
<p>The 47698 fix commit also closes a route where the <code>__proto__</code> setter ran entirely in the host realm through a host array's <code>.map</code>, without crossing the bridge. A check at the bridge cannot see what never crosses it. So 3.11.6 changes approach: bridge functions refuse raw host mutators, a second check rejects host objects with a severed prototype chain, and host errors reach sandbox catch blocks as sandbox-realm errors.</p>
<p>Those are better fixes, and I would still not bet on them being the last. 3.11.4 closed ten advisories and 3.11.6 closed five, under three months apart.</p>
<p>If an agent executes JavaScript a model wrote, treat that code as attacker controlled, because whatever steers the model through a prompt, a document or a tool output also steers the code. Run it in vm2 and you meet the 47698 precondition by design. Check whether your framework does. My position is that the boundary belongs <a class="" href="https://docs.dwaar.ai/blog/the-agent-harness-is-where-ai-starts-acting-security-has-to-start-there-too">in the harness</a> and outside the language runtime: a separate process in a container or VM, with no ambient credentials.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-do-now">What to do now<a href="https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem#what-to-do-now" class="hash-link" aria-label="Direct link to What to do now" title="Direct link to What to do now" translate="no">​</a></h2>
<ol>
<li class=""><strong>Find it.</strong> List vm2 in each project's dependency tree, and check lockfiles and SBOMs for any entry at 3.11.5 or lower, direct or transitive. NVD only carried these from 17 August, so a process keyed on it started three days late.</li>
<li class=""><strong>Upgrade.</strong> Move vm2 to 3.11.6 or later. For a transitive copy, update the parent package or force the version through your package manager. The advisories list no workaround.</li>
<li class=""><strong>Test the behaviour changes.</strong> <code>builtin: ['*']</code> in <code>NodeVM</code> no longer exposes <code>os</code> and <code>dns</code>, a finite <code>bufferAllocLimit</code> now also covers <code>Buffer.concat</code>, <code>Buffer.from</code> and the ArrayBuffer family, and non-primitive properties on caught host errors are gone.</li>
<li class=""><strong>If you cannot upgrade today,</strong> rethrow host errors as plain strings. That is my inference, not official guidance, and it addresses 47686 only. For 47698, stop sending untrusted code to that sandbox.</li>
<li class=""><strong>Hunt.</strong> The sources publish no indicators of compromise. Both PoCs end in a host command, so look for the Node process hosting the sandbox spawning children or writing files it normally never touches.</li>
</ol>
<p>Then take the service where vm2 runs your least trusted code and write down who authors that code, which host functions you pass in, and which credentials and networks the process can reach. If that last list bothers you, move the boundary before the next advisory.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/vm2-sandbox-escapes-and-the-incomplete-fix-problem#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">vm2 (GitHub Security Advisory), <a href="https://github.com/patriksimek/vm2/security/advisories/GHSA-m283-3h24-438v" target="_blank" rel="noopener noreferrer" class="">GHSA-m283-3h24-438v</a>, 2026-08-14</li>
<li class="">vm2 (GitHub Security Advisory), <a href="https://github.com/patriksimek/vm2/security/advisories/GHSA-cfcw-xp6x-25gj" target="_blank" rel="noopener noreferrer" class="">GHSA-cfcw-xp6x-25gj</a>, 2026-08-14</li>
<li class="">NVD, <a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-47686" target="_blank" rel="noopener noreferrer" class="">CVE-2026-47686</a>, 2026-08-17, SSVC entry 2026-08-18</li>
<li class="">NVD, <a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-47698" target="_blank" rel="noopener noreferrer" class="">CVE-2026-47698</a>, 2026-08-17, SSVC entry 2026-08-18</li>
<li class="">vm2 (GitHub release), <a href="https://github.com/patriksimek/vm2/releases/tag/3.11.6" target="_blank" rel="noopener noreferrer" class="">3.11.6</a>, 2026-08-14</li>
<li class="">vm2 (GitHub releases), <a href="https://github.com/patriksimek/vm2/releases/tag/v3.11.4" target="_blank" rel="noopener noreferrer" class="">3.11.4</a> and <a href="https://github.com/patriksimek/vm2/releases/tag/v3.11.5" target="_blank" rel="noopener noreferrer" class="">3.11.5</a>, 2026-05-18</li>
<li class="">vm2 (GitHub commit), <a href="https://github.com/patriksimek/vm2/commit/a85acb61f81402c6eabf32760aa11272af6d0f9e" target="_blank" rel="noopener noreferrer" class="">a85acb61</a>, authored 2026-08-09, released in 3.11.6 on 2026-08-14</li>
<li class="">vm2 (GitHub commit), <a href="https://github.com/patriksimek/vm2/commit/7e3faaf550f4ab975bf4cdde183fcec49b056d8e" target="_blank" rel="noopener noreferrer" class="">7e3faaf5</a>, authored 2026-08-14, released in 3.11.6 on 2026-08-14</li>
<li class="">GitHub Advisory Database, <a href="https://github.com/advisories/GHSA-v6mx-mf47-r5wg" target="_blank" rel="noopener noreferrer" class="">GHSA-v6mx-mf47-r5wg, CVE-2026-47131</a>, 2026-05-18</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>Open source</category>
            <category>AI agents</category>
        </item>
        <item>
            <title><![CDATA[AI-Written Code Changes the Vulnerability Math]]></title>
            <link>https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math</link>
            <guid>https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math</guid>
            <pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Merged pull requests on GitHub rose 29% while issue and PR comments stayed flat, and LLMs invent package names. What that proves, what it does not, and what I would change in the pipeline.]]></description>
            <content:encoded><![CDATA[<p>If your developers use coding assistants, your pull request queue has likely grown, and the number of people qualified to review it probably has not. The usual security response is to argue about whether AI-written code is worse, then add another scanner. I think both moves miss where the pressure is.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-obvious-argument-is-the-weak-one">The obvious argument is the weak one<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#the-obvious-argument-is-the-weak-one" class="hash-link" aria-label="Direct link to The obvious argument is the weak one" title="Direct link to The obvious argument is the weak one" translate="no">​</a></h2>
<p>The obvious argument says AI code carries more vulnerabilities per line, so measure that and gate on it. The evidence is older than it looks. The <a href="https://arxiv.org/abs/2108.09293" target="_blank" rel="noopener noreferrer" class="">2021 Copilot study</a> found about 40% of 1,689 generated programs vulnerable, but it used high-risk CWE scenarios and an early model. The <a href="https://arxiv.org/abs/2211.03622" target="_blank" rel="noopener noreferrer" class="">Stanford user study</a> used Codex-davinci-002. Both are history, not current rates, and no source I have read gives a controlled comparison of vulnerabilities per line between current models and human developers.</p>
<p>Density is also the wrong fight. Hold it constant, double the code, and you have doubled the bugs. Adding a scanner then adds findings to a remediation queue that is already behind. The number I care about is a ratio: what enters your pipeline, over what your people can check and fix.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-terms-in-the-ratio">The terms in the ratio<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#the-terms-in-the-ratio" class="hash-link" aria-label="Direct link to The terms in the ratio" title="Direct link to The terms in the ratio" translate="no">​</a></h2>
<p>What has been measured, and what each number cannot carry.</p>








































<table><thead><tr><th>Term</th><th>What was reported</th><th>What it does not show</th></tr></thead><tbody><tr><td>Code volume</td><td><a href="https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/" target="_blank" rel="noopener noreferrer" class="">GitHub Octoverse 2025</a>: 518.7M merged pull requests (up 29%), more than 986M commits (up 25%)</td><td>That AI caused it. GitHub calls these signals observational</td></tr><tr><td>Review discussion</td><td>Same data: comments on issues and PRs up 0.35%, comments on commits down 27%</td><td>Less review. Comments are a weak proxy for effort</td></tr><tr><td>Invented dependencies</td><td><a href="https://arxiv.org/abs/2406.10279" target="_blank" rel="noopener noreferrer" class="">Spracklen et al.</a>: at least 5.2% of packages for commercial models, 21.7% for open-source models</td><td>A breach. The name still has to be registered and installed</td></tr><tr><td>Human verification</td><td><a href="https://arxiv.org/abs/2211.03622" target="_blank" rel="noopener noreferrer" class="">Perry et al.</a>: assisted participants wrote less secure code and were more likely to believe it was secure</td><td>Current models</td></tr><tr><td>Delivery stability</td><td><a href="https://cloud.google.com/blog/products/devops-sre/announcing-the-2024-dora-report" target="_blank" rel="noopener noreferrer" class="">DORA 2024</a>: estimated 7.2% fall in stability with rising AI adoption. <a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report" target="_blank" rel="noopener noreferrer" class="">DORA 2025</a> still finds a negative relationship with stability, but now a positive one with throughput, which DORA 2024 had estimated falling 1.5%</td><td>Security outcomes. These findings come from survey data</td></tr><tr><td>Patch capacity</td><td><a href="https://helpnetsecurity.com/2026/05/20/verizon-2026-dbir-findings" target="_blank" rel="noopener noreferrer" class="">Verizon DBIR 2026, via Help Net Security</a>: 26% of CISA KEV vulnerabilities fully remediated (38% the year before), median time to full patch 43 days (was 32)</td><td>AI as the cause. Help Net Security notes the analysed data predates the rise in AI-assisted vulnerability research by threat actors</td></tr></tbody></table>
<p>Nothing here proves AI wrote your next CVE. It says something narrower: the input side is growing, and the checking side shows little sign of growing with it.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="dependencies-changed-shape">Dependencies changed shape<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#dependencies-changed-shape" class="hash-link" aria-label="Direct link to Dependencies changed shape" title="Direct link to Dependencies changed shape" translate="no">​</a></h2>
<p>A hallucinated package is not a flaw in code. It is a name nobody owns yet. Spracklen and colleagues tested 16 models across 576,000 Python and JavaScript samples and counted 205,474 unique invented names. <a href="https://arxiv.org/html/2406.10279v3" target="_blank" rel="noopener noreferrer" class="">Repetition is what makes that useful to an attacker</a>: publish a package under a name models keep repeating, and later users who trust the generated code download it.</p>
<p>That is a different failure from a vulnerable dependency, and <a class="" href="https://docs.dwaar.ai/blog/a-vulnerable-package-and-a-malicious-package-are-different-problems">a different problem to solve</a>. A vulnerability scan asks whether a known package has a known flaw. Here the question is whether this package should be in your build at all.</p>
<p>The attack needs a registered name and a person or pipeline that installs it unchecked. The authors report that mitigations reduce hallucination, with RAG and fine-tuning particularly effective, though fine-tuning cost code quality and they still call the problem persistent and systemic. None of the sources behind this post documents a confirmed breach through a registered hallucinated name. The <a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-slopsquatting-ai-supply-chain-20260419/" target="_blank" rel="noopener noreferrer" class="">Cloud Security Alliance</a> does report a malicious npm package under one such name, with downloads but no named victim. I treat this as a cheap hole to close, not a crisis.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="review-is-where-the-honest-uncertainty-sits">Review is where the honest uncertainty sits<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#review-is-where-the-honest-uncertainty-sits" class="hash-link" aria-label="Direct link to Review is where the honest uncertainty sits" title="Direct link to Review is where the honest uncertainty sits" translate="no">​</a></h2>
<p><a href="https://stackoverflow.blog/2025/07/29/developers-remain-willing-but-reluctant-to-use-ai-the-2025-developer-survey-results-are-here" target="_blank" rel="noopener noreferrer" class="">Stack Overflow's 2025 survey</a> has 80% of developers using AI tools while trust in AI accuracy fell from 40% to 29%. That is a self-selected vendor survey, not a security source. It still fits the Stanford result: participants who trusted the AI less and engaged more with their prompts produced fewer vulnerabilities. Scepticism works as a control, and scepticism costs time.</p>
<p>DORA cuts the other way. In its 2024 report, a 25% rise in adoption was associated with 3.4% higher code quality and 3.1% faster code review. So I do not believe AI makes review worse by default. Its <a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report" target="_blank" rel="noopener noreferrer" class="">2025 report</a> reverses the throughput finding but keeps the stability one, and says AI's acceleration can expose weaknesses downstream. My reading: each change can look better and move faster while delivery gets less stable. DORA's own answer is the basics, small batch sizes and robust testing in 2024, strong automated testing and fast feedback loops in 2025, not abandoning the tools.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-queue-it-lands-in">The queue it lands in<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#the-queue-it-lands-in" class="hash-link" aria-label="Direct link to The queue it lands in" title="Direct link to The queue it lands in" translate="no">​</a></h2>
<p><a href="https://www.first.org/blog/20260211-vulnerability-forecast-2026" target="_blank" rel="noopener noreferrer" class="">FIRST's February forecast</a> picked 2026 as the year published CVEs cross 50,000, with a median near 59,000, and called 70,000 to 100,000 realistic. Its <a href="https://www.first.org/blog/20260615-vulnerability-forecast-update" target="_blank" rel="noopener noreferrer" class="">June 15 update</a> reports cumulative drift of 46.3% above the original forecast, an excess of 6,420 CVEs, and revises the 2026 projection to about 66,000. That still sits inside February's 90% interval of 30,012 to 117,673, so read all of it as direction, not a measurement.</p>
<p>FIRST's update attributes part of the surge to AI-assisted bug finding and part to structural CNA expansion. That is a different mechanism from AI-written bugs. It also reports that the subset it treats as actionable, vulnerabilities in CISA KEV or with an EPSS score above 10%, stayed flat in the first half of 2026. It expects the work of patching live systems to stay steady through at least the end of 2026, while telling software maintainers to prepare to double theirs. So on FIRST's evidence the volume is growing and the part that demands urgent patching is not.</p>
<p>Verizon, as reported by Help Net Security, blames slower patching on a drastic increase in the number of vulnerabilities, including those in the KEV catalog. That reads differently from FIRST's flat line, though the windows differ: Help Net Security gives the DBIR's window as incidents from November 2024 to October 2025, and FIRST's overlay covers the first half of 2026. I am not claiming AI-written code produced any of these numbers. I am saying it lands on the same people. FIRST puts the constraint as "the human capacity to verify, coordinate, and patch", and that is the bottom half of my ratio.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-i-would-change">What I would change<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#what-i-would-change" class="hash-link" aria-label="Direct link to What I would change" title="Direct link to What I would change" translate="no">​</a></h2>
<p>Stop trying to review everything harder. Point human attention at the few files where a mistake changes what runs, and let the build refuse what nobody approved.</p>
<p>Start with dependency names. Resolve every new one against an allowlist or internal proxy, and fail the build on a name that was not already in the lockfile and has not been approved. Pin and verify lockfiles in CI. This closes the hallucination path however common the attack turns out to be.</p>
<p>Then require a human reviewer on dependency manifests and CI workflows, regardless of who or what wrote the change. Those diffs are short, and one reviewer buys the most there.</p>
<p>Keep pull requests small, including the ones agents open. A reviewer cannot be sceptical of a diff they cannot read.</p>
<p>Finally, scan the built artifact and rank by <a class="" href="https://docs.dwaar.ai/blog/cvss-is-not-risk-how-to-use-epss-and-kev">KEV and EPSS rather than raw count</a>. If only 26% of known exploited vulnerabilities get fully remediated, a longer list is not what you need, and FIRST's flat KEV and EPSS subset says the new volume is not where the urgent work is. That is the part Dwaar works on: discovering what you ship from source control, registries and cloud accounts, scanning it, and keeping one findings list with EPSS and KEV signals.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="where-this-stops-applying">Where this stops applying<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#where-this-stops-applying" class="hash-link" aria-label="Direct link to Where this stops applying" title="Direct link to Where this stops applying" translate="no">​</a></h2>
<p>If your volume is flat and your dependency tree rarely changes, the ratio has not moved, and none of this is urgent. An allowlist has a real cost for teams that add packages often. The figures also age: the hallucination paper was last revised in March 2025, so your tools may do better or worse.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="try-this-on-your-own-repositories">Try this on your own repositories<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#try-this-on-your-own-repositories" class="hash-link" aria-label="Direct link to Try this on your own repositories" title="Direct link to Try this on your own repositories" translate="no">​</a></h2>
<p>Take your last 90 days of merged pull requests and list every dependency name added to a manifest. For each, record when it was first published, whether anyone but the author reviewed the change, and whether your build would have refused an unapproved name. Count the names no second person looked at. That count tells you more about your exposure than your vulnerability total does.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/ai-written-code-changes-the-vulnerability-math#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class="">Spracklen et al., <a href="https://arxiv.org/abs/2406.10279" target="_blank" rel="noopener noreferrer" class="">We Have a Package for You!</a>, arXiv / USENIX Security 2025, June 12, 2024, revised March 2, 2025 (<a href="https://arxiv.org/html/2406.10279v3" target="_blank" rel="noopener noreferrer" class="">full text</a>)</li>
<li class="">Perry et al., <a href="https://arxiv.org/abs/2211.03622" target="_blank" rel="noopener noreferrer" class="">Do Users Write More Insecure Code with AI Assistants?</a>, arXiv / ACM CCS 2023, November 7, 2022</li>
<li class="">Pearce et al., <a href="https://arxiv.org/abs/2108.09293" target="_blank" rel="noopener noreferrer" class="">Asleep at the Keyboard?</a>, arXiv / IEEE S&amp;P 2022, August 20, 2021</li>
<li class=""><a href="https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/" target="_blank" rel="noopener noreferrer" class="">Octoverse 2025</a>, GitHub Blog, October 28, 2025</li>
<li class=""><a href="https://cloud.google.com/blog/products/devops-sre/announcing-the-2024-dora-report" target="_blank" rel="noopener noreferrer" class="">Announcing the 2024 DORA report</a>, Google Cloud Blog, October 22, 2024</li>
<li class=""><a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report" target="_blank" rel="noopener noreferrer" class="">Announcing the 2025 DORA Report: State of AI-Assisted Software Development</a>, Google Cloud Blog, September 23, 2025</li>
<li class=""><a href="https://stackoverflow.blog/2025/07/29/developers-remain-willing-but-reluctant-to-use-ai-the-2025-developer-survey-results-are-here" target="_blank" rel="noopener noreferrer" class="">2025 Developer Survey results</a>, Stack Overflow Blog, July 29, 2025 (re-published December 29, 2025)</li>
<li class=""><a href="https://www.first.org/blog/20260211-vulnerability-forecast-2026" target="_blank" rel="noopener noreferrer" class="">Vulnerability forecast 2026: The Year Ahead</a>, FIRST, February 11, 2026 (date from the URL; the page is undated)</li>
<li class=""><a href="https://www.first.org/blog/20260615-vulnerability-forecast-update" target="_blank" rel="noopener noreferrer" class="">The 2026 Vulnerability Forecast Update: Navigating the AI Epoch</a>, FIRST, June 15, 2026 (the article says about 66K; the teaser on FIRST's blog index says about 68K)</li>
<li class=""><a href="https://helpnetsecurity.com/2026/05/20/verizon-2026-dbir-findings" target="_blank" rel="noopener noreferrer" class="">Verizon DBIR: Vulnerability exploitation is the dominant initial access vector</a>, Help Net Security, May 20, 2026</li>
<li class=""><a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-slopsquatting-ai-supply-chain-20260419/" target="_blank" rel="noopener noreferrer" class="">Slopsquatting: AI Code Hallucinations Fuel Supply Chain Attacks</a>, Cloud Security Alliance, April 19, 2026</li>
</ul>]]></content:encoded>
            <category>Insights</category>
            <category>AI agents</category>
            <category>Remediation</category>
        </item>
        <item>
            <title><![CDATA[The keyv and cacheable npm Worm: Valid Provenance, Poisoned Source]]></title>
            <link>https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop</link>
            <guid>https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop</guid>
            <pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[A malicious keyv 6.0.0 reached npm through the project's real release workflow with valid provenance, then a worm spread through the Cacheable packages and hundreds more.]]></description>
            <content:encoded><![CDATA[<p>If any machine or build job of yours ran an npm install on or after August 4, you have a question to answer today, even if you never chose keyv: did that install resolve one of the versions below? The malware runs at install time, steals credentials and republishes itself. Its first release carried valid provenance, so an attestation check would have passed it.</p>

































<table><thead><tr><th>Field</th><th>Detail</th></tr></thead><tbody><tr><td>Identifier</td><td>No CVE. <a href="https://github.com/advisories/GHSA-3p9h-f68w-m6fx" target="_blank" rel="noopener noreferrer" class="">GHSA-3p9h-f68w-m6fx</a> and <a href="https://api.osv.dev/v1/vulns/MAL-2026-11524" target="_blank" rel="noopener noreferrer" class="">MAL-2026-11524</a>. Press label: <a href="https://www.infosecurity-magazine.com/news/chaindrop-worm-400-npm-two-billion/" target="_blank" rel="noopener noreferrer" class="">ChainDrop</a></td></tr><tr><td>Affected</td><td>keyv 6.0.0, the Cacheable family, and hundreds of other packages</td></tr><tr><td>Fixed in</td><td>None listed by the advisory, so remove it and return to the prior version. <a href="https://thehackernews.com/2026/08/keyv-linked-npm-worm-poisons-hundreds.html" target="_blank" rel="noopener noreferrer" class="">Reported</a> restored as latest on August 4: keyv 5.6.0, flat-cache 6.1.23, cache-manager 7.2.9</td></tr><tr><td>Severity</td><td>No CVSS, nobody scored it. GitHub classifies it as Malware, CWE-506</td></tr><tr><td>Exploited as of 2026-08-06</td><td>Yes. Self-propagating malware, not a vulnerability</td></tr><tr><td>Disclosed</td><td>2026-08-04</td></tr></tbody></table>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-happened">What happened<a href="https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop#what-happened" class="hash-link" aria-label="Direct link to What happened" title="Direct link to What happened" translate="no">​</a></h2>
<p>A commit titled <code>chore: update config</code> <a href="https://github.com/jaredwray/keyv/commit/d8c850c7800e514d582741a1f852b61122c31213" target="_blank" rel="noopener noreferrer" class="">landed on keyv</a> at 09:04:30 UTC, authored as <code>github-actions[bot]</code> and shown as verified. <a href="https://registry.npmjs.org/keyv" target="_blank" rel="noopener noreferrer" class="">npm timestamps</a> <code>keyv@6.0.0</code> at 09:35:00 UTC. Registry timestamps then show the spread: <a href="https://registry.npmjs.org/cacheable" target="_blank" rel="noopener noreferrer" class="">cacheable 2.5.1</a> at 10:10:44, <a href="https://registry.npmjs.org/flat-cache" target="_blank" rel="noopener noreferrer" class="">flat-cache 6.1.24</a> at 10:10:55, <a href="https://registry.npmjs.org/cache-manager" target="_blank" rel="noopener noreferrer" class="">cache-manager 7.2.10</a> at 10:14:41, <a href="https://registry.npmjs.org/ecto" target="_blank" rel="noopener noreferrer" class="">ecto 5.0.1</a> at 10:28:01.</p>
<p>The counts disagree, and they count artifacts, not victims. <a href="https://thehackernews.com/2026/08/keyv-linked-npm-worm-poisons-hundreds.html" target="_blank" rel="noopener noreferrer" class="">The Hacker News</a> carries 420 names and 1,684 versions from one research team, and <a href="https://csa.gov.sg/alerts-and-advisories/advisories/ad-2026-009" target="_blank" rel="noopener noreferrer" class="">CSA Singapore</a> says over 1,300 versions. <a href="https://www.csoonline.com/article/4205276/chaindrop-credential-stealing-worm-infects-over-400-npm-packages.html" target="_blank" rel="noopener noreferrer" class="">CSO Online</a> quotes 444 packages and 2,212 versions as of 18:10 UTC, with more than 2 billion combined monthly downloads. That measures package reach, not bad installs.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="the-root-cause">The root cause<a href="https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop#the-root-cause" class="hash-link" aria-label="Direct link to The root cause" title="Direct link to The root cause" translate="no">​</a></h2>
<p><strong>Getting in.</strong> <a href="https://www.infosecurity-magazine.com/news/chaindrop-worm-400-npm-two-billion/" target="_blank" rel="noopener noreferrer" class="">Infosecurity Magazine</a> and CSO Online report that the maintainer's GitHub account was compromised. The Hacker News says the initiating credential was not established, and that neither the verified badge nor the bot author identifies who held it. No source says how it happened. CSO paraphrases maintainer Jared Wray as saying he uses OIDC with npm and one-time codes on GitHub.</p>
<p>Infosecurity quotes researchers describing a direct push to main followed by an immediate release. The Hacker News adds that the commit planting the hooks carried a verified badge and a bot author. Either way, the release went through the project's real workflow.</p>
<p><strong>Why the provenance was valid.</strong> That workflow built already trojanized source, and npm attached valid OIDC and SLSA provenance. As the <a href="https://api.osv.dev/v1/vulns/MAL-2026-11524" target="_blank" rel="noopener noreferrer" class="">OSV record</a> puts it, "provenance attests build integrity, not source integrity".</p>
<p><strong>What runs.</strong> <code>package.json</code> gains <code>"preinstall": "node setup.mjs"</code>. The compiled library is untouched, so the package still works. <code>setup.mjs</code> downloads a Bun runtime and uses it to run <code>Math_Symbol.js</code>, an obfuscated bundle of about 728 KB. It goes after cloud metadata endpoints and keys, Vault tokens, Kubernetes service-account tokens, GitHub Actions OIDC and secrets, and npm tokens, then sweeps for other secrets. With the stolen npm identity, including trusted publishing, it republishes whatever that identity can reach. Stolen data is encrypted, and <a href="https://www.elastic.co/security-labs/threat-command/shai-hulud-chaindrop-npm-supply-chain" target="_blank" rel="noopener noreferrer" class="">Elastic Security Labs</a> describes the route out. The worm asks an Ethereum smart contract for the current exfiltration endpoint at runtime, and Elastic observed the dead-drop domains <code>awqhnjewqjkl[.]icu</code> and <code>npm-cache[.]com</code>. If that lookup fails, it searches GitHub commit history for a signed marker that yields a domain. Only if both paths fail does it use a public GitHub repository described as "Shai-Hulud: Here We Go Again".</p>
<p><strong>The path with no install.</strong> Hooks in <code>.claude/settings.json</code> (SessionStart) and <code>.vscode/tasks.json</code> (a task named <code>Environment Setup</code>, run on folder open) call the same loader when a developer or coding agent opens a clone. They fire only if the workspace is trusted or project configuration is permitted. Treat those files as code.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-to-check-and-do-now">What to check and do now<a href="https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop#what-to-check-and-do-now" class="hash-link" aria-label="Direct link to What to check and do now" title="Direct link to What to check and do now" translate="no">​</a></h2>
<p>Search for these exact versions, from the <a href="https://csa.gov.sg/alerts-and-advisories/advisories/ad-2026-009" target="_blank" rel="noopener noreferrer" class="">CSA Singapore advisory</a>. The list is partial:</p>
<ul>
<li class=""><code>keyv</code> 6.0.0, <code>ecto</code> 5.0.1</li>
<li class=""><code>cacheable</code> 2.5.1, <code>cacheable-request</code> 13.0.20, <code>cache-manager</code> 7.2.10</li>
<li class=""><code>flat-cache</code> 6.1.24, <code>file-entry-cache</code> 11.1.6</li>
<li class=""><code>@cacheable/memory</code> 2.2.1, <code>@cacheable/utils</code> 2.5.1, <code>@cacheable/net</code> 2.1.1, <code>@cacheable/node-cache</code> 3.1.2</li>
</ul>
<ol>
<li class=""><strong>Search lockfiles, not <code>package.json</code>, and not <code>latest</code>.</strong> Tags moved during the incident. The Hacker News: "Upgrading alone may therefore preserve the exposure."</li>
<li class=""><strong>Treat a match as a breach, even for an unused transitive dependency.</strong> A preinstall hook does not care whether you import the package. The <a href="https://github.com/advisories/GHSA-3p9h-f68w-m6fx" target="_blank" rel="noopener noreferrer" class="">GitHub advisory</a> says such a machine "should be considered fully compromised". Rebuild it, as CSA advises.</li>
<li class=""><strong>Do not rotate first.</strong> The payload installs a token-revocation watcher, and The Hacker News warns that rotating first "can run an attacker-supplied local handler." Take the host out of service, remove the watcher or rebuild, then rotate from a different, clean machine: cloud, GitHub, SSH, Kubernetes, Terraform and npm credentials.</li>
<li class=""><strong>Hunt the indicators.</strong> Look for an install process fetching Bun 1.3.13 from GitHub releases, the file hashes in the OSV record, <code>.claude/setup.mjs</code> and <code>.vscode/setup.mjs</code>, and npm versions you did not publish. Check network logs for <code>npm-cache[.]com</code> and <code>awqhnjewqjkl[.]icu</code>, and for Bun reaching Ethereum RPC endpoints such as <code>eth-mainnet.nodereal[.]io</code>, because updating the contract rotates the domains. Search for public GitHub repositories with that Shai-Hulud description too, and not only in your organization: <a href="https://insights.integrity360.com/threat-advisories/chaindrop-the-keyv-and-cacheable-npm-supply-chain-attack" target="_blank" rel="noopener noreferrer" class="">one advisory</a> says they are created under compromised identities, so include the personal accounts of your developers and CI identities. Finding no repository does not clear a host, since data may have left over the primary channel. Maintainers should treat publish credentials as exposed.</li>
<li class=""><strong>Change the default.</strong> Use npm's <code>--ignore-scripts</code> in CI and set a minimum release age of 3 to 7 days, which CSO says npm, pnpm, Yarn and Bun support.</li>
</ol>
<p>The Hacker News cites researchers who found the keyv 5.x line and the published <code>@keyv/*</code> adapters clean. Given the reported compromise of the maintainer's account, I would still check any <code>@keyv/*</code> entry in your lockfile against its published tarball.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="what-repeats-and-what-changes">What repeats, and what changes<a href="https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop#what-repeats-and-what-changes" class="hash-link" aria-label="Direct link to What repeats, and what changes" title="Direct link to What repeats, and what changes" translate="no">​</a></h2>
<p>Researchers place this in the Shai-Hulud family, and CSA uses the name. The Hacker News cautions that the overlap "does not establish who operated the August activity". So say same playbook, operator unknown.</p>
<p>The playbook is what stays: steal from developer machines and CI, republish with the stolen identity, and keep public GitHub repositories as one place to drop the loot. The Hacker News reports that researchers documented the same hooks, filename and Bun download in an April compromise of <code>lightning</code> on PyPI. A researcher quoted by Infosecurity says the worm "keeps moving through trusted publishing workflows".</p>
<p>The trigger is what changes. June's <a class="" href="https://docs.dwaar.ai/blog/miasma-npm-worm-red-hat-trusted-publishing">Miasma packages</a> <a href="https://thehackernews.com/2026/06/miasma-supply-chain-attack-compromises.html" target="_blank" rel="noopener noreferrer" class="">used a preinstall hook</a> and wrote a <code>SessionStart</code> hook and a <code>folderOpen</code> task as persistence. Three weeks ago, <a class="" href="https://docs.dwaar.ai/blog/asyncapi-pwn-request-trojaned-packages-with-valid-provenance">trojaned AsyncAPI packages</a> also shipped with valid provenance, but <a href="https://thehackernews.com/2026/07/compromised-asyncapi-npm-packages.html" target="_blank" rel="noopener noreferrer" class="">with no install script</a>: the payload ran when the module loaded. This one is back to a preinstall hook, with that editor and agent config planted in the repository as a second door. Detection pinned to one trigger ages fast, so build on what recurs: what a build job's identity can read and publish, and what a tarball does when installed or loaded.</p>
<p>Keep verifying provenance, but know where it stops. Reading the tarball is what shows a preinstall hook calling a file the library never references. And with no CVE and no CVSS, a queue sorted by severity has nothing to sort.</p>
<p>Pick one job that installed npm packages on August 4. List the exact versions it resolved, then every secret it could read. Shrink that second list this week, whether or not the first one has a match.</p>
<h2 class="anchor anchorTargetStickyNavbar_Vzrq" id="references">References<a href="https://docs.dwaar.ai/blog/keyv-cacheable-npm-worm-chaindrop#references" class="hash-link" aria-label="Direct link to References" title="Direct link to References" translate="no">​</a></h2>
<ul>
<li class=""><a href="https://github.com/advisories/GHSA-3p9h-f68w-m6fx" target="_blank" rel="noopener noreferrer" class="">GHSA-3p9h-f68w-m6fx</a>, GitHub Advisory Database, 2026-08-04</li>
<li class=""><a href="https://api.osv.dev/v1/vulns/MAL-2026-11524" target="_blank" rel="noopener noreferrer" class="">MAL-2026-11524</a>, OSV / OpenSSF malicious-packages, 2026-08-04</li>
<li class=""><a href="https://registry.npmjs.org/keyv" target="_blank" rel="noopener noreferrer" class="">keyv package metadata</a>, npm registry, version timestamps, 2026-08-04</li>
<li class=""><a href="https://registry.npmjs.org/cacheable" target="_blank" rel="noopener noreferrer" class="">cacheable package metadata</a>, npm registry, version timestamps, 2026-08-04</li>
<li class=""><a href="https://registry.npmjs.org/flat-cache" target="_blank" rel="noopener noreferrer" class="">flat-cache package metadata</a>, npm registry, version timestamps, 2026-08-04</li>
<li class=""><a href="https://registry.npmjs.org/cache-manager" target="_blank" rel="noopener noreferrer" class="">cache-manager package metadata</a>, npm registry, version timestamps, 2026-08-04</li>
<li class=""><a href="https://registry.npmjs.org/ecto" target="_blank" rel="noopener noreferrer" class="">ecto package metadata</a>, npm registry, version timestamps, 2026-08-04</li>
<li class=""><a href="https://github.com/jaredwray/keyv/commit/d8c850c7800e514d582741a1f852b61122c31213" target="_blank" rel="noopener noreferrer" class="">Commit d8c850c7800e</a>, GitHub, 2026-08-04</li>
<li class=""><a href="https://thehackernews.com/2026/08/keyv-linked-npm-worm-poisons-hundreds.html" target="_blank" rel="noopener noreferrer" class="">Keyv-linked npm worm poisons hundreds</a>, The Hacker News, 2026-08-04</li>
<li class=""><a href="https://www.csoonline.com/article/4205276/chaindrop-credential-stealing-worm-infects-over-400-npm-packages.html" target="_blank" rel="noopener noreferrer" class="">ChainDrop credential-stealing worm infects over 400 npm packages</a>, CSO Online, 2026-08-04</li>
<li class=""><a href="https://www.infosecurity-magazine.com/news/chaindrop-worm-400-npm-two-billion/" target="_blank" rel="noopener noreferrer" class="">ChainDrop worm, 400+ npm packages</a>, Infosecurity Magazine, 2026-08-05</li>
<li class=""><a href="https://insights.integrity360.com/threat-advisories/chaindrop-the-keyv-and-cacheable-npm-supply-chain-attack" target="_blank" rel="noopener noreferrer" class="">ChainDrop: The Keyv and Cacheable npm Supply Chain Attack</a>, Integrity360, 2026-08-05</li>
<li class=""><a href="https://csa.gov.sg/alerts-and-advisories/advisories/ad-2026-009" target="_blank" rel="noopener noreferrer" class="">Ongoing npm Supply Chain Attack Affecting Keyv and Related Packages</a>, Cyber Security Agency of Singapore, 2026-08-06</li>
<li class=""><a href="https://www.elastic.co/security-labs/threat-command/shai-hulud-chaindrop-npm-supply-chain" target="_blank" rel="noopener noreferrer" class="">Shai-Hulud strikes again: CHAINDROP worm hits 400+ npm packages</a>, Elastic Security Labs, 2026-08-06</li>
<li class=""><a href="https://thehackernews.com/2026/06/miasma-supply-chain-attack-compromises.html" target="_blank" rel="noopener noreferrer" class="">Miasma Supply Chain Attack Compromises Red Hat npm Packages with Credential-Stealing Worm</a>, The Hacker News, 2026-06-01</li>
<li class=""><a href="https://thehackernews.com/2026/07/compromised-asyncapi-npm-packages.html" target="_blank" rel="noopener noreferrer" class="">Compromised AsyncAPI npm Packages Deliver Multi-Stage Botnet Malware</a>, The Hacker News, 2026-07-15</li>
</ul>]]></content:encoded>
            <category>Incidents</category>
            <category>Supply chain</category>
            <category>Open source</category>
        </item>
    </channel>
</rss>