audit-loop
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill establishes a structured 'Audit Loop' protocol that prioritizes security through role separation. It explicitly mandates that auditor sub-agents (e.g., oracle, metis) must be read-only and prohibits them from executing fixes or modifying files.
- [INDIRECT_PROMPT_INJECTION]: The skill acknowledges the risk of processing potentially untrusted data (like audit objects or git logs). It mitigates this by requiring a 'six-segment' delegation structure (TASK, EXPECTED OUTCOME, etc.) that clearly defines boundaries and constraints for the sub-agents, reducing the likelihood of data-driven prompt injection influencing the auditor's behavior.
- [COMMAND_EXECUTION]: The skill references standard development tools such as
git logandtypecheckfor verifying fixes. These are used appropriately within the context of a development-focused AI agent and do not involve arbitrary or unauthorized command execution.
Audit Metadata