frame
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data such as user ideas, context, and external documents, which creates a theoretical surface for indirect prompt injection attacks.
- Ingestion points: The agent is instructed to read supplied ideas, context, and external retrieved material (webpages, documents, repository files) as defined in the discovery protocol.
- Boundary markers: The skill contains robust mitigation in the 'Untrusted-source hard boundary' section of its protocol, which explicitly mandates that external material cannot change the user's goal or grant unauthorized permissions.
- Capability inventory: The skill has the capability to write records to the local filesystem (e.g., under '.itp/work/') and suggests handoffs to tools like brainstorming.
- Sanitization: The instructions explicitly forbid the agent from executing commands embedded in source material and require SHA-256 hash verification for proposed specifications before they are considered approved.
Audit Metadata