evaluate
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFE
Full Analysis
- [DATA_EXPOSURE]: The skill uses the 'opik' SDK to transmit experiment data, traces, and metrics to the Comet ML platform. This is the core functionality of the skill and uses the vendor's official infrastructure.
- [INDIRECT_PROMPT_INJECTION]: The skill defines evaluation workflows where LLMs (judges) process untrusted data (model traces). While this presents a surface for indirect prompt injection, the skill provides extensive documentation on mitigating this through specific, binary evaluation criteria and structured outputs.
- Ingestion points: Evaluation prompts in 'references/write-judge-prompt.md' and 'references/evaluate-rag.md' process user queries and model outputs.
- Boundary markers: Templates use markdown headers and structured sections to separate instructions from data.
- Capability inventory: The system is designed to score and log results; no sensitive system-level capabilities are exposed to the judge's output.
- Sanitization: The instructions emphasize human review of traces and validation of judges against expert labels to ensure system integrity.
Audit Metadata