Posts
High-signal AI/security/automation notes.
HTTP/2 Bomb — OpenAI Codex Chains Decade-Old DoS Into Web Server Crash
JFrog — IronWorm: Rust-Based npm Worm with eBPF Rootkit Targets AI/Dev Tooling
JFrog discovers IronWorm, a Rust-built self-replicating npm worm with eBPF kernel rootkit that harvests AI developer credentials (Codex, Anthropic, Gemini, Cursor) and propagates via GitHub.
Kiteworks — The Lethal Trifecta: Why Most AI Agents Are Structurally Exploitable
Kiteworks research identifies a lethal trifecta — any AI agent that accesses private data, processes untrusted content, and can communicate externally is structurally exploitable, confirmed by documented production failures.
Miasma Worm Targets AI Coding Agents via GitHub Repo Config Injection
A parallel arm of the Miasma worm campaign pushes malicious commits to 120+ GitHub repos, exploiting Claude Code, Cursor, and VS Code auto-run features to detonate a credential harvester on folder open.
Microsoft — Updated Agentic AI Failure Modes Taxonomy (7 New)
Microsoft expanded its agentic AI failure modes taxonomy with seven new categories covering supply chain, goal hijacking, CUA visual attacks, and MCP/plugin abuse.
Adversa SymJack — Symlink-Hijack RCE in AI Coding Agents
VIPER-MCP Scans 40,000 Repos, Finds 106 Zero-Days in MCP Servers
VIPER-MCP framework scanned ~40,000 open-source MCP server repositories and discovered 106 zero-day taint-style vulnerabilities, producing 67 CVEs to date.
Anthropic Frontier Red Team — AI Cyber Threats Mapped to MITRE ATT&CK
CVE-2026-45497 — Microsoft 365 Copilot Critical Command Injection RCE
Microsoft disclosed CVE-2026-45497, a CVSS 9.8 critical RCE via command injection in the M365 Copilot orchestration engine, silently patched in the cloud.
GitGuardian MCP Governance Framework for Enterprise AI
GitGuardian publishes a practical MCP governance framework covering authentication standards, scope control, secrets lifecycle, and credential exposure detection for multi-agent deployments.
Exposed MCP Servers Triple to 1,467; Trend Micro Finds Cloud Command-Injection Flaws
Three independent scans in June 2026 show ~40% of remote MCP servers run without authentication, while Trend Micro found CVSS 9.8 command-injection flaws in unofficial cloud MCP implementations.
Meta AI Support Bot — Instagram Account Takeover via Prompt Injection
AIRQ Report — Only 11% of Production Agents Pass AI Security Bar
An independent assessment of 100 production AI agents finds only 11% pass basic security, with coding and computer-use agents carrying the highest risk.
Anthropic Opus 4.8 Browser Agent — 31.5% Pre-Safeguard Hijack Rate
Claroty Team82 — LLMs Discover Critical ICS Vulnerabilities Without Prior Disclosures
Claroty Team82 uses Claude Opus 4.6 to rediscover five critical ICS vulnerabilities on a Zenitel intercom platform without access to prior vulnerability details.
CVE-2026-27735 — MCP Reference Servers git_add Path Traversal
CVE-2026-27735 in MCP reference servers mcp-server-git allows path traversal via the git_add tool, staging files outside repository boundaries.
DesktopCommanderMCP SSRF (CVE-2026-10690) — MCP File Tool Exposes IMDS
Four Major AI Labs Use Incompatible Prompt-Injection Metrics
Anthropic, OpenAI, Google, and Meta each publish prompt-injection benchmarks in 2026, but no two labs measure the same threat — leaving buyers with no common baseline for model safety claims.
LibreChat CVE-2026-32625 — MCP URL Env-Var Interpolation Leaks Secrets
LibreChat ≤0.8.3 interpolates ${VAR} placeholders in user-supplied MCP server URLs against process.env, exfiltrating CREDS_KEY, JWT_SECRET, and MONGO_URI to attacker-controlled hosts.