arena
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill requires the agent to read and evaluate content produced by multiple sub-agents in Phase D ('Read every candidate end to end before picking'). Since this content is generated by other LLM instances, it could potentially contain instructions that influence the parent agent's behavior. The skill lacks specific boundary markers or instructions to disregard potential prompts within the candidate artifacts.
- [COMMAND_EXECUTION]: The skill orchestrates the creation of sub-agents using platform-specific tools like 'Agent' (Claude Code) and 'spawn_agent' (Codex). It also involves file system operations such as creating 'git worktree' instances or writing to temporary directories (e.g., '/tmp/arena-slug/') to manage candidate outputs.
Audit Metadata