verify
Pass
Audited by Gen Agent Trust Hub on Sep 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted trajectory traces which may contain embedded instructions designed to subvert the agent's judging logic or influence ranking outcomes.
- Ingestion points: External trajectories are ingested during the generation phase in SKILL.md and processed for evaluation in the pairwise prompt template defined in REFERENCE.md.
- Boundary markers: The skill specifies the use of clear delimiters (e.g., '=== TRACE_A START ===' and '=== TRACE_A END ===') to isolate untrusted content from the rest of the prompt.
- Capability inventory: The agent is tasked with ranking these traces and may be required to execute empirical checks (code execution) within isolated file system scopes like worktrees or copies.
- Sanitization: The instructions provide explicit guidance to the agent to treat all fence-bounded text as untrusted data, to ignore any narration or instructions found within those fences, and to strip HTML comments from criteria files.
Audit Metadata