write-agent-guidance
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data to generate authoritative agent instructions, creating a vulnerability surface where malicious instructions embedded in project files could be promoted to the agent's core instruction set.\n
- Ingestion points: Repository manifests, task scripts, CI workflows, architecture documents, and source files (SKILL.md Section 2).\n
- Boundary markers: The instructions recommend ordering content by consequence and using direct imperative language, but do not provide explicit delimiters or "ignore" instructions for the processed content.\n
- Capability inventory: File write operations for AGENTS.md and CLAUDE.md files which are automatically loaded by agents.\n
- Sanitization: The skill includes logic to explicitly exclude secrets, PII, and generic advice, though it does not specifically filter for embedded prompt injection patterns in the source artifacts.\n- [EXTERNAL_DOWNLOADS]: The skill references external documentation and research papers.\n
- Evidence: Links to official documentation on developers.openai.com and claude.dev, as well as research papers on arxiv.org.\n
- Note: These references target well-known and reputable organizations and research services.
Audit Metadata