▣ Skill Poison — Knowledge ExplorerONLINE / EDITION 10
AGENT SKILLS / SECURITY
Independent. Open. Curious.
THE ARCHIVE / EDITION 08

Reading trails

Five routes through the archive, for newcomers, reviewers, researchers, and maintainers.

These are editorial reading paths, not certifications or required courses. Start anywhere; every guide keeps its sources attached.
READING TRAIL

01 / A first visit

Understand the package before studying the attacks.

  1. What is skill poisoning?Start here · 5 min
  2. The anatomy of an agent skillBeginner · 3 min
  3. Instructions, data, and authorityIntermediate · 3 min
  4. What marketplace studies can actually tell usIntermediate · 2 min
READING TRAIL

02 / At the review desk

From a package identity to a bounded decision.

  1. Threat-model a skill before you install itIntermediate · 3 min
  2. A repeatable skill admission reviewIntermediate · 3 min
  3. Least privilege for installed skillsTechnical · 3 min
  4. The skill poisoning defense checklistPractical · 3 min
  5. From scanner warning to defensible findingPractical · 2 min
  6. Poisoned advice and output integrityIntermediate · 2 min
  7. Provenance, pinning, and updatesIntermediate · 3 min
  8. Respond to a suspected poisoned skillPractical · 3 min
READING TRAIL

03 / Follow the evidence

Read a result with its version, denominator, and limits.

  1. How to read skill poisoning researchIntermediate · 3 min
  2. Read the version, not just the headlineBeginner · 2 min
  3. Why a scanner can look strong and still failTechnical · 2 min
  4. What model-internal detectors can and cannot showTechnical · 2 min
  5. A clean result with missing coverageTechnical · 2 min
  6. Evaluate the complete skill lifecycleAdvanced · 3 min
READING TRAIL

04 / For skill authors

Make the package and its resulting behavior easier to inspect.

  1. From a description to an actionIntermediate · 2 min
  2. When documentation becomes executable outputTechnical · 2 min
  3. Scripts and dependencies inside skillsTechnical · 3 min
  4. The prerequisite trap: when setup becomes the attackPractical · 2 min
  5. Publish skills that are easier to trustFor maintainers · 3 min
READING TRAIL

05 / The next research frontier

Skills can carry models, interact, and emerge from experience.

  1. Backdoors in model-bearing skillsAdvanced · 3 min
  2. When skills combine: composition riskAdvanced · 3 min
  3. When one skill supplies the pretext for anotherAdvanced · 2 min
  4. When experience becomes a poisoned skillAdvanced · 3 min

Need a place for your own evidence? Open the review notebook. It stays in your browser.

Search guides, research, and glossary termsESC TO CLOSE

Start typing to explore the knowledge base.