evaluate
Pass
Audited by Gen Agent Trust Hub on Jun 19, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No security issues detected. The skill's functionality is consistent with its documentation, providing a robust utility for model performance assessment.\n- [PROMPT_INJECTION]: The skill identifies a standard attack surface for indirect prompt injection when evaluating LLM outputs.\n
- Ingestion points: Untrusted model outputs from predictions.csv and results.tsv are interpolated into LLM-as-judge evaluation prompts (e.g., PAIRWISE_PROMPT in SKILL.md).\n
- Boundary markers: The prompt templates utilize Markdown headers and triple quotes as delimiters to structure the input, though they do not include explicit instructions to the judge LLM to disregard potentially malicious instructions within the data.\n
- Capability inventory: The skill leverages tool access (Bash, Read, Write, Glob, Grep) to manage and analyze evaluation data.\n
- Sanitization: Input data is interpolated directly into prompts without formal escaping; however, the skill implements adversarial robustness through position-bias mitigation logic (swapping A/B positions and checking for consistency).
Audit Metadata