human-intervention
Pass
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns, obfuscation, or unauthorized data access were detected. The skill instructions are focused on process management and documentation.
- [PROMPT_INJECTION]: The skill manages untrusted human feedback, which represents an indirect prompt injection surface. This risk is effectively mitigated by design through the separation of raw feedback from agent interpretation and the use of explicit boundary markers.
- Ingestion points: Human input is collected in human-interventions/active/*/content.md files.
- Boundary markers: The templates use blockquotes (">") to delimit the human's verbatim words from the rest of the document.
- Capability inventory: Actions are restricted to file system operations (creation, modification, and movement of files) within the project's documentation structure.
- Sanitization: The protocol enforces an "Agent Interpretation" and "Impact Assessment" phase, ensuring that the agent evaluates the input before formalizing it into an action plan.
Audit Metadata