product-critic-acrazie
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a 'Project loading boundary' that prevents the agent from reading or executing dependencies without explicit user commands. It explicitly forbids editing code, installing dependencies, or running implementation tasks.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted repository data (code, tests, manifests) in SKILL.md. This risk is mitigated by explicit boundary markers requiring read-only inspection, prohibiting code execution, and limiting capabilities to writing reports in 'docs/critics/' only after human approval. The instructions also specify sanitizing secrets from reports.
- [SAFE]: The suggested dependency installation ('npx skills add Acrazie/skills@interview-acrazie') targets a vendor-owned resource ('acrazie') and is presented as a proposal for the user to initiate, rather than a command the agent executes silently.
Audit Metadata