llm-evaluation

Pass

Audited by Gen Agent Trust Hub on Aug 6, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: Indirect prompt injection vulnerability surface detected in the LLM-as-Judge implementations in reference/evaluation-harness.md and reference/trajectory-evaluation.md.\n
  • Ingestion points: Untrusted agent responses and user queries are ingested into the evaluation context via the llm_judge and EffectivenessJudge components.\n
  • Boundary markers: The judging prompts use clear labels for inputs but lack strict boundary delimiters or instructions to ignore embedded commands within the content being evaluated.\n
  • Capability inventory: Since the results are intended for use in CI/CD gates and automated benchmarks, successful injection could lead to misleading quality metrics or bypassed safety checks.\n
  • Sanitization: The skill does not perform sanitization or escaping on the content being evaluated before it is interpolated into the judge's prompt.\n- [EXTERNAL_DOWNLOADS]: The skill facilitates the download of standard data and models from well-known services.\n
  • Downloads text processing data from the NLTK registry using nltk.download in reference/evaluation-metrics.md.\n
  • Fetches pre-trained machine learning models for NLI and toxicity evaluation from the Hugging Face model hub.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 6, 2026, 01:53 PM
Security Audit — agent-trust-hub — llm-evaluation