evaluation-harness
Warn
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: MEDIUMPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION] (MEDIUM): The
score_with_llmfunction inSKILL.mdimplements an 'LLM as judge' pattern that is vulnerable to indirect prompt injection. - Ingestion points: It processes untrusted model outputs (
actual) and dataset entries (expected) which are interpolated into the prompt. - Boundary markers: No delimiters or isolation techniques are used in the prompt template to separate data from instructions.
- Capability inventory: The resulting scores can trigger
sys.exit(1), potentially failing automated builds and affecting the deployment process. - Sanitization: The input is interpolated directly into the judge's prompt without any sanitization or validation of its content.
Audit Metadata