iterate-until-verified
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates on user-supplied tasks, creating an attack surface for indirect prompt injection. The design includes mitigation strategies:
- Ingestion points: Arbitrary task descriptions are ingested in SKILL.md for extraction and decomposition.
- Boundary markers: The instructions mandate the use of an 'AUTHORITATIVE TASK' block to isolate the user request.
- Capability inventory: The workflow mentions delegating to subagents or running validation tools, but these are scoped by the platform's existing permissions.
- Sanitization: Explicit commands prevent the agent from relaxing security constraints or expanding permissions based on the input.
- [SAFE]: All external references point to well-known technology platforms and no suspicious code or obfuscation is present.
Audit Metadata