Posts
High-signal AI/security/automation notes.
Filter: All AI Vulnerabilities & CVEs (312)Tools & Comparisons (2)AI Security Research (208)Agent & Supply Chain Security (699)
VentureBeat — CLI-Anything AI Agent Skill Backdoor and Structural Supply-Chain Gap
CLI-Anything auto-generates AI agent skill files (SKILL.md) that no SAST or SCA scanner can detect, creating a new supply-chain attack surface across all major coding agents.
FastGPT — SSRF Cluster in AI Agent Platform
Four SSRF-related CVEs disclosed in FastGPT, an open-source AI agent platform: unauthenticated SSRF, MCP tool URL bypass, DNS rebinding, and inconsistent SSRF protection.
Forcepoint — 10 In-the-Wild Indirect Prompt Injection Payloads Targeting AI Agents
Forcepoint researchers discovered 10 real-world indirect prompt injection payloads hidden in web content, targeting AI agents for financial fraud, data destruction, and API key theft.
Jeff Kaufman — AI Is Breaking Two Vulnerability Disclosure Cultures
MCPwn — First Named MCP Exploit Campaigns with Actively Exploited CVEs
Pluto Security disclosed MCPwn (CVE-2026-33032, CVSS 9.8, actively exploited, 2,600+ instances exposed) and MCPwnfluence (CVE-2026-27825/27826, unauthenticated RCE via SSRF + arbitrary file write in mcp-atlassian).
Noma Security — MCP Servers and Skills: The Observability Gap in AI Agent Deployments
Noma Security analyzed hundreds of MCP servers and Skills: one in four MCP servers expose agents to code execution risk, and real-world attack chains like ContextCrush and ForcedLeak are already active.
Vercel Breached via Context AI Supply Chain — Internal Database Sold for $2M
Anthropic — Natural Language Autoencoders Reveal Models Can Detect Safety Tests
EU AI Act Simplification — Nudification Ban Enacted, High-Risk Rules Delayed
European lawmakers agreed to ban AI nudification tools while delaying high-risk AI enforcement to December 2027, reshaping the AI Act compliance timeline.
SentinelLabs — PCPJack Cloud Worm Steals AI API Keys, Evicts TeamPCP
A new cloud-focused malware framework called PCPJack scans for exposed services and steals credentials including OpenAI and Anthropic API keys, while actively removing rival TeamPCP infections.
Elastic Security — TCLBanker Banking Trojan Distributed via Trojanized AI Prompt Builder
A new banking trojan called TCLBanker uses a trojanized Logitech AI Prompt Builder MSI installer to infect systems, demonstrating how AI tool branding is weaponized for initial access.
SecurityWeek — Critical Ollama CVE-2026-7482 Exposes 300K Deployments
Sophos — Fake Claude AI Website Delivers Beagle Windows Backdoor
A fake Claude AI website distributes a trojanized installer that deploys the Beagle backdoor, targeting developers who work with AI coding tools.
Keep Aware — Browser DLP Blind Spot Lets Sensitive Data Leak to AI Prompts
arXiv — MOSAIC-Bench: Coding Agents Ship Exploitable Code via Sequenced Prompts
New benchmark shows nine production coding agents from Anthropic, OpenAI, and Google compose innocuous tickets into exploitable code at 53–86% success rate, bypassing per-prompt safety reviews.
pnpm 11 — Default Supply Chain Protections Against Malicious Package Installs
pnpm 11 enforces a 24-hour minimum release age for new packages by default, reducing risk from typosquatting and supply-chain attacks that target automated installs.
vm2 — CVE-2026-26956 Sandbox Escape with Published PoC
Critical sandbox escape in the vm2 Node.js library (1.3M+ weekly downloads) enables arbitrary host code execution via WebAssembly exception handling bypass; PoC exploit published.
arXiv — Agentic Red Teaming Agent Compresses AI Testing from Weeks to Hours
New arXiv paper introduces an agentic red teaming system built on the Dreadnode SDK that automates attack selection, transform composition, and reporting — achieving 85% attack success against Llama Scout with zero human-written code.
arXiv — Tool-Mediated LLM Architecture with Lean 4 Proofs for Autonomous Cyber Defense
New paper presents a tool-mediated LLM architecture for autonomous SOC operations, with machine-checked Lean 4 proofs guaranteeing stability — Claude Sonnet 4 reduces attacker payoff by 59% on 282 real enterprise attack graphs.