eval-harness

Pass

Audited by Gen Agent Trust Hub on Mar 24, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill framework describes the use of shell commands (e.g., bash, npm test, grep) to perform deterministic code-based grading. These operations are intended for local environment testing and project verification as part of the evaluation workflow.
  • [PROMPT_INJECTION]: The skill introduces a 'Model-Based Grader' pattern where the agent evaluates output quality. This presents an indirect prompt injection surface where the content being evaluated could attempt to influence the agent's judging logic.
  • Ingestion points: Evaluation definitions in .claude/evals/*.md and code changes submitted for grading.
  • Boundary markers: Uses clear Markdown headers for grader prompts, though it does not explicitly define delimiters for untrusted evaluated content.
  • Capability inventory: The skill uses Read, Write, Edit, Bash, Grep, and Glob tools to manage the lifecycle of evaluations.
  • Sanitization: No automated sanitization of external content is specified; however, the skill mandates 'Human review for security' as a best practice to mitigate automated evaluation risks.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 24, 2026, 07:41 AM
Security Audit — agent-trust-hub — eval-harness