Security Review for Third-Party Skills

Classify the skill's risk level by inspecting its frontmatter and directories, then apply the matching review tier. Doc-level skills get a light review of author, license, and content; compute and network skills require a formal code audit and sandbox test execution.

Use pip-audit for Python requirements, npm audit for Node.js packages, or Snyk for general dependency scanning. The review requires pinned dependency versions from trusted registries like PyPI or the official npm registry.

Check the SKILL.md frontmatter for network flags and dependencies, then look for scripts directories. Skills with network calls are network-level, scripts without network are compute-level, and markdown-only skills are doc-level.

Red flags include eval or exec on user input, reading SSH keys or browser credentials, exfiltrating environment variables, hardcoded API keys, path traversal outside the skill directory, and obfuscated or minified code.

Run the script in an isolated Docker container, monitor network traffic with tcpdump or Wireshark, and diff file system snapshots before and after execution. Verify the observed behavior matches the skill's documentation.

Reject skills with any code audit red flags, suspicious network calls to undisclosed endpoints, obfuscated code, or credential harvesting patterns. Minor concerns can warrant conditional approval with trusted set to false and ongoing monitoring.