arena
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The workflow requires an agent to read multiple generated outputs to synthesize a final result, creating a risk that malicious instructions within those outputs could influence the process.
- Ingestion points: Candidate artifacts and rationales are processed during the Cross-judge and Pick phases as defined in
references/arena-procedure.md. - Boundary markers: The instructions do not establish clear delimiters or directives to ignore instructions embedded within the processed data.
- Capability inventory: The skill enables spawning subagents and executing shell commands for environment isolation.
- Sanitization: No sanitization of external or generated content is documented before processing.
- [COMMAND_EXECUTION]: The skill uses shell commands for workspace management and agent orchestration.
- Evidence:
SKILL.mdandreferences/arena-procedure.mdmention utilizinggit worktreefor isolated checkouts and spawning "background launcher processes" for external lanes.
Audit Metadata