eval-harness
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The evaluation workflow requires the agent to ingest and process untrusted content such as code changes and eval definitions. This content could contain malicious instructions designed to bypass constraints or manipulate the evaluation results. 1. Ingestion points: File content read from the repository for capability and regression testing (e.g., .claude/evals/). 2. Boundary markers: The skill does not define specific delimiters or instructions to prevent the agent from obeying instructions embedded in the test data. 3. Capability inventory: The skill has access to powerful tools including Bash, Write, Edit, and Grep. 4. Sanitization: There is no evidence of content sanitization before processing eval outputs or grader results.
- [COMMAND_EXECUTION]: The framework relies on executing shell commands via the Bash tool for deterministic grading. Examples include npm test, npm run build, and node scripts/eval-harness.js. These operations run with the permissions of the agent and interact directly with the project environment.
Audit Metadata