evaluation-harness

Pass

Audited by Gen Agent Trust Hub on Sep 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The score_with_llm function creates a vulnerability by interpolating untrusted model output directly into a prompt for a judge LLM.
  • Ingestion points: Model output (actual) generated during evaluation cycles and expected responses from the dataset are used as input for the judge.
  • Boundary markers: Absent. The prompt template does not use delimiters (such as XML tags or unique separators) to differentiate the model output from the judge's instructions.
  • Capability inventory: The scores returned by the judge are parsed and used to decide CI/CD outcomes, potentially allowing malicious model output to bypass quality gates.
  • Sanitization: No escaping or filtering mechanisms are implemented for model outputs before they are processed by the evaluation logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 4, 2026, 09:30 AM
Security Audit — agent-trust-hub — evaluation-harness