The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from repository files, issues, pull requests, and web pages, creating a surface for indirect prompt injection. This risk is effectively mitigated by an explicit 'Untrusted Content Boundary' defined in SKILL.md and references/safety.md. These instructions mandate that the agent treat all data as evidence rather than instructions, ignore any embedded commands, and refrain from executing or installing audited content. The capability inventory is strictly limited to read-only search and inspection tasks, and the skill enforces data sanitization by requiring the agent to extract facts only and quote minimally.\n- [SAFE]: The skill is designed to be read-only, and the documentation explicitly forbids state mutations or repository writes without separate, explicit user approval. All provided validation and testing scripts rely exclusively on the Python standard library, avoiding the use of external packages or remote code downloads. No patterns of data exfiltration or unauthorized network communication were identified during analysis.

pre-build-feature-audit