principle-never-block-on-the-human
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill encourages reduced human-in-the-loop (HITL) oversight by instructing the agent to 'Proceed, then present' and 'reserve confirmation for irreversible actions'.
- Ingestion points: This instruction influences how the agent handles all tasks, including processing external data or files that may contain untrusted instructions.
- Boundary markers: The skill includes a 'Boundaries' section that limits autonomy to reversible actions (e.g., writing code) and mandates confirmation for high-impact actions (e.g., deleting data, sending messages), providing a safety buffer.
- Capability inventory: The instruction applies to all file-writing and task-execution capabilities available to the agent.
- Sanitization: By prioritizing 'async supervision' and execution-before-review, the agent may act on malicious data before a human can intervene, although the reversible nature of the work allows for later correction.
Audit Metadata