llm-evaluation

Pass

Audited by Gen Agent Trust Hub on Mar 28, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill's implementation of the 'LLM-as-Judge' pattern is vulnerable to indirect prompt injection because it interpolates untrusted data (the model outputs being evaluated) directly into the judge's prompt template.
  • Ingestion points: Model outputs are ingested via the response, response_a, and response_b parameters in functions llm_judge_quality and compare_responses within SKILL.md.
  • Boundary markers: Absent. The code uses standard f-string interpolation without delimiters (such as XML tags or triple backticks) or instructions for the judge to ignore embedded instructions.
  • Capability inventory: The evaluated content is passed to openai.ChatCompletion.create. No shell execution or persistent file system operations are triggered by this process.
  • Sanitization: The skill does not perform any sanitization, escaping, or filtering of the model responses before placing them in the prompt.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 28, 2026, 08:17 PM
Security Audit — agent-trust-hub — llm-evaluation