An independent research library on skill poisoning: how reusable AI agent skills gain authority, how that trust can fail, and what the evidence supports.
Welcome, reader. Pick a thread. Follow the sources. Stay curious.
Fill in the checklist and your findings, then download your review as JSON or Markdown. Your entries stay in this browser.
Manual review worksheet · 12 checks · Local autosaveA journal article proposes pre-integration review of MCP tool manifests and skill files, using structural analysis and light dynamic checks.
The revised PERSIST study explores a skill that satisfies the visible user task while carrying an attacker-desired extra action in a separate imported module.
SkillRelayBench evaluates how attacker-controlled skills can coordinate through artifacts that survive a session reset.
A complete starting point: what gets poisoned, how a skill gains influence, and which boundaries determine the outcome.
A practical sequence for deciding what to approve, what evidence to retain, and when to keep a skill out of a sensitive workflow.
Review installation requests and human handoffs as carefully as the actions an agent executes itself.
Follow evidence, state, and authority across a skill chain when no single file explains the final action.
Read detection research in terms of observed signals, benchmark conditions, and deployment limits.
How to write a review result that does not confuse a heuristic, a trace, an attempted action, and impact.
Follow a link. Find a new idea. This independent web archive connects skill security explanations with original research. Choose a folder, save a guide, or follow the RSS feed to return later.
About this archive →