eval-harness-builder
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The 'Judge Prompt Template' defined in Step 3 is vulnerable to indirect prompt injection. It interpolates untrusted data from multiple sources into a higher-privileged evaluator prompt.
- Ingestion points: The template in
SKILL.mduses placeholders{query},{expected_answer},{actual_response}, and{context_if_applicable}to build the judge's context. - Boundary markers: The template uses basic markdown formatting (headers and bold text) to separate fields, but lacks explicit instructions for the judge model to ignore any commands or malicious directives contained within the
{actual_response}variable. - Capability inventory: The skill orchestrates LLM generation calls as part of the evaluation pipeline.
- Sanitization: There is no evidence of sanitization, escaping, or validation logic to process the inputs before they are interpolated into the judge prompt.
Audit Metadata