arena
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to spawn multiple parallel processes and subagents described as 'direct background launcher processes' and 'background subagents' in Phase B (SKILL.md).
- [INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow where an agent processes potentially untrusted content generated by other model instances.
- Ingestion points: The parent agent and a dedicated 'judge' agent ingest artifacts and rationales produced by candidate runners during the 'Cross-judge', 'Pick a base', and 'Graft' phases (SKILL.md).
- Boundary markers: The instructions do not specify the use of delimiters or 'ignore embedded instructions' warnings when the agent or judge reads the outputs of the candidates.
- Capability inventory: The skill environment allows for spawning subagents, executing background processes, and performing file system operations in
/tmpand git worktrees (SKILL.md). - Sanitization: There is no mention of sanitizing, escaping, or validating the model-generated rationales or artifacts before they are used to influence the final synthesis and verification steps.
Audit Metadata