gs-arena
Pass
Audited by Gen Agent Trust Hub on Nov 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator that takes user-defined tasks and interpolates them into prompts for multiple subagents. This creates a surface where instructions embedded in the user's task could potentially influence the subagents' behavior.
- Ingestion points: The skill ingests untrusted user tasks and context in the 'Frame' and 'Fan out' phases to define the artifact and rubric.
- Boundary markers: The skill explicitly instructs that the prompt is the 'contract' and candidates see only the task, which provides a conceptual boundary for the subagents.
- Capability inventory: The skill uses the harness's subagent tools for parallel processing and interacts with the file system to create git worktrees or temporary directories for each candidate.
- Sanitization: There is no explicit sanitization or escaping of the user-provided content before it is passed into the candidate prompts or the cross-judge prompt.
Audit Metadata