▣ Skill Poison — Knowledge ExplorerONLINE / EDITION 10
AGENT SKILLS / SECURITY
Independent. Open. Curious.
← THE GUIDE LIBRARY
DefensePractical

The skill poisoning defense checklist

A concrete review baseline across package content, authority, execution, updates, and incident readiness.

THE THREE THINGS TO KEEP
  • Review content, code, and authority together.
  • A checked box should have evidence behind it.
  • Make revocation part of readiness.

Package identity

  • Record source, maintainer, version or commit, and artifact integrity.
  • Assign an internal owner and intended task scope.
  • Inventory all files, resources, dependency sources, and learned artifacts.
  • Document unresolved provenance questions rather than silently accepting them.

Instruction and content review

  • Compare the description and body with the actual task.
  • Follow every reference and inspect how assets are consumed.
  • Investigate unrelated prerequisites, concealment requests, and changed destinations.
  • Separate task procedure from permission to perform sensitive actions.
  • Record mutable remote resources and their change policy.

Code and runtime

  • Review executable entry points, subprocesses, and dependencies.
  • Test with synthetic records and no production credentials.
  • Limit readable and writable paths to the workflow.
  • Restrict network destinations and inspect data leaving the environment.
  • Validate output-to-action paths, especially for model-bearing components.
  • Test the complete composed workflow, not only each package alone.

Approval and observation

  • Enforce sensitive actions outside the skill’s own instructions.
  • Show the actual target and meaningful arguments in an approval.
  • Log package identity, actions, approvals, and effects with sensitive values protected.
  • Distinguish static warnings, attempted actions, and confirmed impact.
  • Keep a known way to disable a revision and revoke associated credentials.

Changes and maintenance

  • Pin the approved artifact and dependency state.
  • Review content, code, resource, permission, and environment changes.
  • Retain an approved rollback artifact.
  • Remove unused skills and unneeded scopes.
  • Re-run boundary tests after material changes.

Use the workbench for a review record and export it when complete. The checklist is an editorial baseline; adapt it to the system and keep evidence behind each decision.

Turn a checked item into a verifiable claim

WORKED EXAMPLE / SYNTHETIC, NOT A REPORTED INCIDENT

A reviewer checks “network access defined.” Another person cannot tell whether that means the skill documentation mentioned a domain, the runtime enforced an allowlist, or a denied transfer was actually tested. The check is a summary of a claim, not the evidence that supports it.

Work through the evidence

For important checks, attach an artifact, observation, and owner. A file-inventory check points to a manifest; a permissions check points to the enforced profile; a test check points to input, expected result, action trace, and outcome. Record skipped checks and explain what blocks completion. Do not invent a reassuring percentage from the number of checked boxes.

Decision criteria and a useful review record

Close a check when its evidence answers the stated question for the exact reviewed state. “Not applicable” also needs a reason: absence of a bundled script does not establish that no external execution path exists. Use the checklist as a handoff between reviewers, and retain the unresolved items beside the final scope decision.

Sources & scope

This guide combines editorial explanation and practical review suggestions. Linked sources support the underlying concepts; this site has not independently replicated their experiments.

S02
Equipping agents for the real world with Agent SkillsDocumentation · Anthropic Engineering
↗
S08
Agent Skill Security: Threat Models, Attacks, Defenses, and EvaluationPreprint · Sanket Badhe and Priyanka Tiwari
↗
S10
Making Claude Code more secure and autonomous with sandboxingDocumentation · Anthropic Engineering
↗
Compare sources in the catalog ↗
NEXT IN THE READING PATHDetection and testing for poisoned skills
Search guides, research, and glossary termsESC TO CLOSE

Start typing to explore the knowledge base.