ps-arena
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes artifacts and rationales generated by multiple parallel subagents. Maliciously crafted content within a subagent's output could potentially influence the selection (Phase D) or grafting (Phase E) process.
- Ingestion points: Phase C (Cross-judge) and Phase D (Pick a base) involve reading candidate artifacts and rationales produced in Phase B.
- Boundary markers: The skill relies on a framing contract defined in Phase A, but does not specify delimiters or 'ignore embedded instructions' warnings for processing candidate-generated data.
- Capability inventory: Phase B involves spawning subagents (run_in_background: true) and assigning output paths.
- Sanitization: No explicit sanitization or validation of the candidate-generated content is mentioned before it is read by the judge or the parent agent.
Audit Metadata