Dwaar Blog
November 2026

When an Attack Costs $25, Compliance Is Not a Security Program
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.

Trusting a Leading Agentic Harness Does Not Guarantee the Security of Your Application
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.

Reachability Analysis: Why Most Dependency CVEs Don't Matter, and How to Prove It
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.
October 2026

Do You Know Which AI Agents You Run?
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.

Stop Counting Vulnerabilities
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.

php2shell: Two Unauthenticated PHP RCEs, One Image Rebuild per Service
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.

CVE-2026-97229: Unauthenticated Code Execution in GitLab Through a Crafted CI/CD Regular Expression
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.

Argo CD Repo Server Command Execution Through Kustomize
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.

A Waiver Should Expire
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.

CVE-2026-94603: A Podman Image Annotation Switches Off the Sandbox You Asked For
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.
September 2026

MCP SDK OAuth Clients Could Send Credentials to a Server-Chosen Authorization Server
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.

CVE-2026-94545: One Satori Escaping Bug, Moderate in the Library and Critical in next/og
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.

The APIs Nobody Told You About
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.

The EU Cyber Resilience Act Reporting Clock Starts This Week
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.

A Finding Is a Claim. Exploitability Is the Proof.
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.
August 2026

ContextForge CVE-2026-53710 and Chainlit CVE-2026-45018: Allowlists and Denylists Around MCP Keep Failing
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.

The Case for a Dependency Cooldown
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.

vm2 Again: Sandbox Escapes and the Incomplete Fix Problem
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.

AI-Written Code Changes the Vulnerability Math
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.

The keyv and cacheable npm Worm: Valid Provenance, Poisoned Source
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.

The 100 BTC Dare Misses the Real Agent-Security Lesson
Why the 100 BTC dare matters less than the security boundary around AI agents and how to control egress, tools, approvals and runtime actions.
July 2026

CVE-2026-66066: Rails Active Storage File Read Through libvips
A crafted upload may read server files through libvips in Rails Active Storage. Fixed in 7.2.3.2, 8.0.5.1 and 8.1.3.1, which need libvips 8.13 or later. Rotate your secrets too.

wp2shell: An Unauthenticated RCE Chain in WordPress Core
Two WordPress core bugs chain into unauthenticated RCE. Fixes shipped July 17, CISA listed both on July 21, and WPScan and CISA score the two CVEs in opposite order.

The Hugging Face Breach Is a Preview: Why Guardrails Must Live Outside the Model
Hugging Face's July 2026 incident shows what happens when the model supply chain itself is compromised. Scope-based identity isn't enough — agentic security needs zero-trust, intent-based identity and a physically separate guardrail engine.

MCP Security Checklist for Agent Builders
A practical checklist for securing MCP-connected AI agents, covering tool access, permissions, third-party servers, runtime controls, and trust boundaries.

AsyncAPI: Trojaned npm Packages With Valid Provenance
Five malicious @asyncapi npm versions shipped on 2026-07-14 through the project's own release pipeline, with valid provenance and no install script.

The Agent Harness Is Where AI Starts Acting. Security Has to Start There Too.
AI agent harnesses turn models into operational systems. Learn where agentic trust boundaries fail and how to secure inputs, memory, tools, and actions.

Secrets Live in More Places Than Your Repository
A deleted secret can survive in git history and image layers, and it keeps working until it is rotated. Why removal is not the fix, and how to test your own rotation.

GitLost: When the Agent's Token Scope Decides What a Public Issue Can Steal
Researchers showed a public issue steering a GitHub Agentic Workflow into posting a private repo's README as a public comment. No CVE. GitHub's gh-aw project began shipping a mitigation on July 8.

The Vulnerability Debt You Inherit From Your Base Image
An image is a snapshot, so what sits under FROM ages from the day you build it. What studies found about inherited vulnerabilities, and what fewer packages, rebuilds and digest pins each buy.
June 2026

How Does dwaar Red Teaming Work?
dwaar red teaming helps teams test AI agents before deployment by mapping workflows, running adversarial attacks, identifying vulnerabilities, and generating reproducible security findings.

Cordyceps: Exploitable GitHub Actions Workflows in Over 300 Repositories
Researchers report more than 300 exploitable repositories from a class of GitHub Actions flaws, including chains of individually valid workflows. No CVE, no versions to patch, no in-the-wild use reported.

Your CI Pipeline Is Production
CI jobs hold publish tokens, cloud credentials and repository write access. Moved tags, pull_request_target, shared caches and stored secrets have all turned that authority on users.

Why Agentic Identity Is the Missing Layer in LLM Security?
For decades, enterprise security has been built around a simple question: who is the user, and what are they allowed to access? Identity and Access Management has solved ...

Whitepaper: Compromising an Agentic QA Platform — A Black-Box Red-Team Walkthrough
How dwaar.ai used layered reconnaissance and multi-turn prompt injection to turn an AI test-failure analyzer into a remote-code-execution...

A vLLM Dockerfile Dependency Confusion: How One --extra-index-url Put the Official Image at Risk (GHSA-jrf6-vqxq-pjv2)
vLLM's Dockerfile mixed --extra-index-url with uv's unsafe-best-match, exposing vllm/vllm-openai builds below 0.22.1 to dependency confusion. The advisory reports no exploitation.

Enterprise AI Doesn’t Need Another Dashboard. It Needs an Intent-Aware Policy Plane
Why Enterprise AI Needs an Intent-Aware Policy Plane? Learn why logs and visibility alone cannot secure enterprise AI—and how adversarial testing, intent-aware policies, and runtime enforcement can keep agents within their authorized boundaries.

Miasma: How a Compromised Red Hat GitHub Account Published Malware Through npm Trusted Publishing
Red Hat says a compromised GitHub account pushed unauthorized commits. Reports say an added workflow then published malicious @redhat-cloud-services packages through npm trusted publishing.

How Is AI Red Teaming Different from VAPT?
AI red teaming and VAPT test different security risks. Learn how AI red teaming goes beyond traditional VAPT to test prompts, context, tools, guardrails, and agent behavior.
May 2026

What an SBOM Tells You, and What It Does Not
An SBOM is an inventory, not a risk assessment. What it answers, where generated SBOMs go wrong, what VEX adds, and where each of them stops.

Why Cloud Posture Is Not Enough
Posture checks show which cloud settings are wrong, not what runs there or whether it can be exploited. Google's cloud data and two broader incident reports rank software exploitation first.

Nx Console 18.95.0: A Trojaned VS Code Extension Published With Stolen Credentials
Nx Console 18.95.0 was trojaned on 18 May 2026 using a developer's leaked GitHub credentials. What it stole, how to check your machines, and what GitHub has and has not said.

NGINX Rift, CVE-2026-42945: An 18-Year-Old Heap Overflow in the Rewrite Module
A heap overflow in the NGINX rewrite module, fixed on 13 May 2026, needs a specific rewrite pattern to trigger. Check versions, images and config.

Securing Tool Graph demands adaptive guardrails
The real risk is no longer just that a model may produce an unsafe answer. The bigger risk is that an agent may take an unsafe action — using real tools, permissions, data, and business workflows. That is why the Model Context Protocol...

The TanStack npm Compromise: Actions Cache Poisoning and a Stolen OIDC Token Behind a Legitimate Trusted Publisher
On 2026-05-11, 84 malicious versions of 42 @tanstack npm packages shipped through TanStack's real trusted publisher, using a poisoned Actions cache and an OIDC token read from runner memory.

Bleeding Llama, CVE-2026-7482: An Unauthenticated Heap Leak in Ollama
An out-of-bounds read in Ollama before 0.17.1 lets an unauthenticated client push heap memory to its own registry. Upgrade, then find every server you run and who can reach it.

Why Existing Security Controls Are Not Enough for Agentic Applications
Model guardrails and cloud controls are useful, but agentic applications need prompt-based red teaming, evidence-backed posture reports, and runtime policy enforcement.

Trajectory-Aware Evolutionary Search
Why the future of AI agent security is about trajectories, not just prompts.
April 2026

Copy Fail, CVE-2026-31431: A Page-Cache Write From a Container to Root
A logic flaw in the Linux kernel's AF_ALG path lets an unprivileged local user write into the page cache of any readable file and gain root. Patch the host kernel, not the image.

A Vulnerable Package and a Malicious Package Are Different Problems
The OSV and GitHub records for the poisoned axios and LiteLLM releases list no CVE. Malicious packages need a different signal, clock and response from vulnerable ones.

Code Scanning Is Not Enough: Scan the Artifact You Ship
Repository scans miss base image packages, build-time downloads, bundled binaries and secrets in layers. Artifact scans have their own blind spots, so run both and gate on the artifact.

MCP STDIO: When Command Execution Is the Design
MCP's stdio transport launches a configured command. Researchers tie that to flaws in tools such as Flowise, fixed in 3.1.0, and Agent Zero 0.9.8. They say the underlying flaw remains.

Agentic AI Security Market Worth $13.52 Billion by 2032
The Agentic AI Security market is projected to grow from $1.65B in 2026 to $13.52B by 2032.

You Cannot Secure What You Cannot List
CISA's BOD 23-01 has agencies run automated discovery every 7 days and scan what it finds. The case for applying that logic to source control, registries and cloud accounts, and where it stops.

CVE-2026-34197: A 13-Year-Old ActiveMQ Bug, Found With AI Assistance
An ActiveMQ Classic management API lets console users run code on the broker. Fixed in 5.19.4 and 6.2.3. The finder says the bug sat for 13 years and was found with Claude.

The axios npm Hijack: When the Attack Goes Through the Maintainer
Two axios releases on npm carried a RAT through a postinstall hook in a new dependency, with no CVE and no change to axios source. What to check, and what lockfiles and release age buy you.
March 2026

Why Remediation Has to Move at Agentic Speed
Mandiant estimates mean time to exploit at minus 7 days, while edge-device fixes reportedly take a median 32 days. Which patching steps agents can take over, and which stay with a human.

The LiteLLM PyPI Compromise: How Stolen CI Credentials Reportedly Reached an AI Gateway
litellm 1.82.7 and 1.82.8 on PyPI carried a credential stealer, which the maintainers believe traces to a publishing token exposed in CI. What to check, rotate and pin.

ingress-nginx CVE-2026-3288 and CVE-2026-4342: Config Injection as the Project Retires
Two ingress-nginx advisories describe nginx config injection via Ingress objects, with cluster-wide Secret access by default, as maintenance ends. Upgrade to 1.13.9, 1.14.5 or 1.15.1.

CVE-2026-33017: Langflow's Public Flow Endpoint Runs Unauthenticated Code, and No Stable Release Fixes It
An unauthenticated Langflow RCE has a public advisory and proof of concept, and as of March 19 no stable langflow release on PyPI carries the fix.

CVSS Is Not Risk: How to Use EPSS and KEV
CVSS Base measures severity, KEV records confirmed exploitation and EPSS forecasts it. How to combine the three into a queue, and where all of them stop.

CVE-2026-29000: Forging pac4j-jwt Logins With Only the Public Key
pac4j-jwt skipped signature checks on unsigned tokens inside encrypted JWEs, so an RSA public key could forge a login. Upgrade to 4.5.9, 5.7.9 or 6.3.3.

Why Scanning Alone Is Not a Security Program
A scan report is a list, not a decision. Using NIST guidance, Verizon's 2025 DBIR and the Equifax record, this post lays out what has to sit between the two.

