2026-07-28NEW
Agent sandbox escapes, CI/CD agents, CoT Forgery
- New Agent sandbox escape section - DuneSlide (CVE-2026-50548/50549, CVSS 9.8, Cato AI Labs): read-only content reaching OS-level RCE through an LLM-controlled
working_directory and a symlink canonicalisation bypass.
- New Agents in CI/CD section - the five-link chain RyotaK used against
claude-code-action (CVE-2025-66032): [bot] actor bypass, injection via issue body, /proc/self/environ, OIDC token exchange, supply-chain spread.
- Chain-of-thought spoofing now cites CoT Forgery (ICML 2026) with its measured ~60% success rate, and the role-confusion reason it transfers across models.
2026-07-14
Agent memory poisoning + DeepTeam
- New Agent memory poisoning section - retrieval-triggered backdoors that persist across sessions (MemPoison, MINJA), plus cross-site memory attacks on AI browsers (ChatGPT Atlas, Perplexity Comet).
- Added
DeepTeam (Confident AI) to Testing & Tooling - open-source agent red-teaming mapped to the OWASP Agentic (ASI) Top 10.
2026-06-30
Arcanum Prompt Injection Taxonomy integrated
- New Reasoning-Model Attacks section - CoT hijacking, thinking-mode steering, chain-of-thought spoofing, structured-output coercion.
- New jailbreaks in Modern Jailbreaks - Echo Chamber, Bad Likert Judge, Deceptive Delight, past/future-tense, plus Self-Persuasion, DarkCite, SATA and AutoDAN-Turbo.
- Six new tricks in Encoding & Hiding - Adversarial Poetry, MathPrompt, QueryAttack, CodeAttack, Trojan Source (bidi) and layered encoding chains.
- New agent-ecosystem attacks in the Agentic Frontier - Tool Rug Pull, Tool Squatting, Tool-Call Spoofing, Rules-File Backdoor, Prompt Worm and Sleeper payloads.
- New Arcanum PIT Cross-Reference table mapping the taxonomy's codes to this site.
- Eleven new Terminology entries covering all of the above.
2026-06-22
PayloadsAllTheThings batch
- Added Code Execution, Memory & Plugin Chaining - RCE in code interpreters, persistent memory injection (spAIware), cross-plugin request forgery and metadata/comment injection.
- Threaded the new techniques across Methodology, the Attack Flow tree and Sources.
- Four new glossary terms and a set of titled academic-paper references.
2026-06-20
Reading & navigation upgrades
- Dimmed reading mode and in-app text zoom (A- / A+, or the + - 0 keys), both remembered between visits.
- Shareable URL hashes for every tab and section, plus hover anchor links on headings.
- rez0's methodology broken out as its own card; performance pass and Lighthouse CI added.