caveman-search
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a multi-agent research workflow that includes an optional verification pass to cross-check claims and correct potential inaccuracies.\n- [SAFE]: Includes a specific safety valve (Auto-Clarity Exception) that overrides stylistic constraints for security warnings or irreversible actions, ensuring critical information is communicated clearly and without ambiguity.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from web research and local repository contents. While this involves ingesting untrusted content, the risk of following embedded instructions is mitigated by the mandatory stylistic transformation (caveman voice) and the optional adversarial verification step.\n
- Ingestion points: Step 3 (repo mode and web mode research agents).\n
- Boundary markers: Not explicitly defined for input, but the voice spec transform serves as a functional filter during rendering.\n
- Capability inventory: Execution of parallel research agents and verifiers.\n
- Sanitization: Use of a secondary verification agent, the Honesty Rule for preventing tool hallucinations, and the safety-focused Auto-Clarity Exception.
Audit Metadata