eval-harness
Pass
Audited by Gen Agent Trust Hub on Jul 4, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The framework processes evaluation definitions and grader instructions from markdown files stored in the
.claude/evals/directory. This establishes a surface for indirect prompt injection where malicious instructions embedded in project files could influence agent behavior. - Ingestion points: Reads evaluation criteria and model-based grader prompts from files in the
.claude/evals/path (SKILL.md). - Boundary markers: The framework does not define the use of delimiters or instructions to ignore embedded commands when processing these external files.
- Capability inventory: The skill uses tools including Bash, Read, Write, and Edit for running tests and managing evaluation history.
- Sanitization: No sanitization or verification logic is provided for the content of the ingested evaluation files.
Audit Metadata