eval-harness
Pass
Audited by Gen Agent Trust Hub on Aug 12, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides templates for using the
Bashtool to run local commands such asnpm testandnpm run buildto verify code functionality deterministically. - [PROMPT_INJECTION]: The framework's 'Model-Based Grader' feature introduces an indirect prompt injection surface by passing untrusted code changes or agent outputs to the model for qualitative evaluation.
- Ingestion points: Project code, model outputs, and file content processed by the
[MODEL GRADER PROMPT]template. - Boundary markers: The grader prompts do not include explicit delimiters or instructions to ignore commands embedded within the evaluated data.
- Capability inventory: The skill has access to tools including
Read,Write,Edit,Bash,Grep, andGlob. - Sanitization: There are no procedures for sanitizing or filtering external content before it is processed by the model-based grader.
Audit Metadata