arena
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill involves reading and processing outputs from multiple parallel LLM candidates. This creates an ingestion surface where a candidate's output could potentially contain instructions aimed at the parent agent. However, the skill includes explicit safety mitigations, such as the 'Verify' phase and the instruction to 'Graft' content manually ('Don't paste mechanically') rather than simple automated interpolation.
- [COMMAND_EXECUTION]: The workflow describes launching background processes and native subagents to perform tasks. This process-spawning behavior is the core intended functionality of the 'Arena' fan-out strategy and relies on the agent's pre-configured provider dispatch environment.
- [SAFE]: The skill uses standard local file paths for staging data (such as
/tmp/and git worktrees) and references local configuration files for model routing, which is consistent with legitimate development environment workflows.
Audit Metadata