eval-harness
Pass
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONNO_CODE
Full Analysis
- [PROMPT_INJECTION]: The skill defines an evaluation workflow that executes tasks described in project-local markdown files.
- Ingestion points: The agent is instructed to read evaluation definitions and task descriptions from the .claude/evals/ directory (e.g., feature-xyz.md).
- Boundary markers: The skill lacks explicit instructions for the agent to use delimiters or to ignore potential system-level instructions embedded within these files.
- Capability inventory: The agent uses Bash, Write, Edit, and Read tools to fulfill the tasks described in the evaluation files.
- Sanitization: There are no requirements for sanitizing or validating the content of the evaluation files before the agent processes them.
- Mitigation: Wrap external or evaluation content in distinct delimiters and include clear instructions to ignore any embedded agent directives.
- [COMMAND_EXECUTION]: The framework utilizes bash commands such as grep and npm test to perform deterministic code-based grading of task completion.
- [NO_CODE]: No executable scripts or binary files were found accompanying this skill; it consists primarily of instructional documentation.
Audit Metadata