harness-setup
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes project-specific data (e.g., plan.md, architecture reports) to generate persistent instruction files such as CLAUDE.md and AGENTS.md. This ingestion of untrusted local data represents a surface for indirect prompt injection.
- Ingestion points: Reads project documentation files in Phase A (Stack Detection) and Phase B (Feedforward Guide Setup).
- Boundary markers: Missing explicit delimiters when interpolating data from project reports, although it references an external audit protocol.
- Capability inventory: The skill can modify project-wide instruction files and invoke other review-oriented skills.
- Sanitization: Employs mandatory human review via the 'ask user question tool' before finalizing any changes to the instruction guides, effectively mitigating the risk of automated injection.
Audit Metadata