arena
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses Git commands (
git worktree add,git worktree remove) to manage isolated environments for different model candidates during the evaluation process. - [INDIRECT_PROMPT_INJECTION]: The skill operates by passing user tasks to multiple sub-agents and then having a 'judge' agent and the main agent review their outputs. This multi-step pipeline creates a surface where output generated by one model (responding to potentially untrusted task input) is later processed as instructions or data by other agents.
- Ingestion points: User-provided task descriptions in Phase A and candidate outputs in Phase B are processed by subsequent agent roles.
- Boundary markers: No specific delimiters or safety warnings for sub-agent prompts are mentioned in the instructions to prevent obedience to embedded instructions.
- Capability inventory: The skill performs file system writes to temporary directories and git worktrees, executes Git commands, and includes a verification phase that may involve running tests or commands on the generated artifacts.
- Sanitization: No explicit sanitization or validation of the candidates' outputs is described before they are read by the judge or parent agent.
Audit Metadata