Skip to main content

Dwaar Blog

November 2026

October 2026

Insights · Oct 29, 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.

Insights · Oct 27, 2026

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.

Incidents · Oct 22, 2026

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.

Incidents · Oct 14, 2026

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.

Incidents · Oct 8, 2026

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.

Insights · Oct 6, 2026

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.

Incidents · Oct 1, 2026

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

August 2026

Incidents · Aug 26, 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.

Insights · Aug 20, 2026

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.

Incidents · Aug 18, 2026

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.

Insights · Aug 12, 2026

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.

Incidents · Aug 6, 2026

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.

Incidents · Aug 5, 2026

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

Incidents · Jul 31, 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.

Incidents · Jul 22, 2026

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.

Incidents · Jul 21, 2026

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.

Insights · Jul 20, 2026

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.

Incidents · Jul 16, 2026

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.

Insights · Jul 13, 2026

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.

Insights · Jul 10, 2026

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.

Incidents · Jul 8, 2026

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.

Insights · Jul 2, 2026

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

Insights · Jun 29, 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.

Incidents · Jun 25, 2026

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.

Insights · Jun 18, 2026

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.

Insights · Jun 16, 2026

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 ...

Insights · Jun 15, 2026

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...

Incidents · Jun 11, 2026

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.

Insights · Jun 10, 2026

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.

Incidents · Jun 4, 2026

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.

Insights · Jun 1, 2026

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

Insights · May 28, 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.

Insights · May 26, 2026

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.

Incidents · May 20, 2026

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.

Incidents · May 15, 2026

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.

Insights · May 14, 2026

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...

Incidents · May 12, 2026

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.

Incidents · May 7, 2026

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.

Insights · May 5, 2026

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.

Insights · May 1, 2026

Trajectory-Aware Evolutionary Search

Why the future of AI agent security is about trajectories, not just prompts.

April 2026

Incidents · Apr 30, 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.

Insights · Apr 29, 2026

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.

Insights · Apr 21, 2026

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.

Incidents · Apr 17, 2026

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.

Insights · Apr 15, 2026

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.

Insights · Apr 10, 2026

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.

Incidents · Apr 8, 2026

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.

Incidents · Apr 2, 2026

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

Insights · Mar 31, 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.

Incidents · Mar 26, 2026

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.

Incidents · Mar 20, 2026

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.

Incidents · Mar 19, 2026

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.

Insights · Mar 11, 2026

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.

Incidents · Mar 6, 2026

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.

Insights · Mar 3, 2026

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.