audit-evals
Pass
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes project-specific data from
regression/cell-history.json. The bundled scriptscripts/audit_evals.pyimplements strict schema validation and type checking to ensure the data adheres to a specific versioned format, effectively mitigating risks of malicious payloads being reflected in audit reports. - Ingestion points:
regression/cell-history.jsonis read and parsed by the Python audit script. - Boundary markers: Structured JSON parsing and version verification act as boundaries.
- Capability inventory: The skill can read history files, write markdown reports, and update the
tierfield in the history JSON. - Sanitization: The script validates that result strings contain only 'p' (pass) or 'f' (fail) characters and ensures suite/case/check identifiers do not contain path delimiters.
- [EXTERNAL_DOWNLOADS]: The skill mentions an installation command for a secondary tool (
llm-evals) provided by the same author. This is presented as a prerequisite check for the user to address manually, ensuring transparency and user control over external software installation.
Audit Metadata