eval-harness
Pass
Audited by Gen Agent Trust Hub on Sep 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes the Bash tool to run deterministic code graders, such as npm test and grep, to verify implementation behavior against success criteria. These are standard operations for a development-focused evaluation harness.
- [DYNAMIC_EXECUTION]: The framework references local mechanical utilities in scripts/eval-harness.js. Importantly, the documentation specifies that candidate execution is globally disabled (gate.isolation_required) until a verified OS containment backend is implemented, preventing the execution of potentially malicious code during evaluation.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-defined evaluation instructions from the .claude/evals/ directory. 1. Ingestion points: .claude/evals/ directory. 2. Boundary markers: Absent in markdown templates. 3. Capability inventory: Read, Write, Edit, Bash, Grep, Glob. 4. Sanitization: Arbitrary candidate execution is explicitly disabled, and the framework mandates human review for security-critical tasks.
Audit Metadata