startup-design
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection due to its heavy reliance on processing untrusted external data.
- Ingestion points: External data enters the agent context through web search results in Phase 0.5 (Pre-Flight Check) and the four waves of Phase 3 (Market Research), as well as through user-provided data in Phase 3.7 (Customer Discovery interviews).
- Boundary markers: The skill lacks explicit delimiters for external content; however, it mandates metadata tagging (e.g.,
[Data],[Estimate],[Assumption]) as defined in thehonesty-protocol.md, which creates some structural separation. - Capability inventory: The skill uses the
Agenttool to spawn sub-tasks,WebSearchfor data collection, and extensive file-system writes to create project documentation. - Sanitization: The process includes a mandatory
Verification Agentpass (Phase 3.5a) described inverification-agent.md, which audits all generated deliverables for inconsistencies, unlabeled claims, and data gaps before the user proceeds.
Audit Metadata