llm-judge-alignment
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns, prompt injections, or unauthorized command executions were found. The skill operates as a purely instructional framework for analyzing evaluator performance metrics.
- [EXTERNAL_DOWNLOADS]: Recommends the installation of the 'judgy' Python package via pip. This is a legitimate tool used for estimating success rates and confidence intervals in the context of LLM evaluation.
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process human-labeled examples and production traces. While this presents an ingestion surface for untrusted data, the agent's actions are limited to classification and analysis with no risky capabilities (such as network writes or file modifications) triggered by the input data.
- [DATA_EXFILTRATION]: No sensitive file access or unauthorized data transfer patterns were detected. References to project context files like 'CLAUDE.md' are standard for development workflows.
Audit Metadata