arena
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates a process where a parent agent spawns multiple subagents (candidates) and then reads their outputs 'end to end' for judging and merging. This creates a surface where adversarial instructions within a candidate's output could influence the parent agent's logic.
- Ingestion points: SKILL.md Phase D ('Read every candidate end to end before picking').
- Boundary markers: Absent. No delimiters or 'ignore embedded instructions' warnings are specified for the evaluation of candidate artifacts.
- Capability inventory: The skill orchestrates multiple agents via
/swarmand performs file operations viagitand/tmppaths. - Sanitization: Absent. The candidate output is read and processed directly without verification or filtering.
- [COMMAND_EXECUTION]: The skill relies on shell-based tools and internal framework commands for environment setup and agent management.
- Evidence: Uses
git worktreefor branch-based isolation and the/swarmcommand for parallel subagent execution.
Audit Metadata