The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute local Python, Bash, and Node scripts for auditing and verifying target AI agent skills. It includes mitigations such as running with shell=False, using isolated temporary home directories, and requiring an explicit trust confirmation flag (--confirm-trusted-candidate) before executing any scripts found within the target folders.
[EXTERNAL_DOWNLOADS]: The security_scan.py script queries the Open Source Vulnerabilities (OSV) database at api.osv.dev to check for known vulnerabilities in the dependencies of analyzed skills. This is a well-known and trusted service provided by Google for security auditing.
[INDIRECT_PROMPT_INJECTION]:
- Ingestion points: The skill accepts a path to an external 'candidate' skill folder via command-line arguments in security_scan.py, harness_evidence_gate.py, and instruction_stability_gate.py.
- Boundary markers: The skill's instructions and implementation require the user to explicitly acknowledge the risks of running untrusted code by providing a confirmation flag.
- Capability inventory: The skill can execute arbitrary local commands via subprocess.run, read and write files within the analyzed and output directories, and perform network requests to the OSV API.
- Sanitization: The safe_path and safe_resolve functions are used to validate file paths, ensuring operations stay within the candidate or policy roots and preventing path traversal attacks.

skill-lint