omh-repro-audit
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute the 'omh' command-line tool for importing records ('omh paper import-harness'), initializing cases ('omh paper audit-init'), and running the audit process ('omh paper audit').- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted evaluation logs (Harness results JSON), creating a potential attack surface for instructions embedded within the data.
- Ingestion points: Reads Harness results JSON and local normalized case files.
- Boundary markers: The skill references the use of 'validation.json' and reporting sanitized reason codes instead of raw source contents.
- Capability inventory: Uses the 'omh' CLI for auditing but explicitly prohibits model execution, subprocess spawning from the audit feature, and network requests.
- Sanitization: Implements validation logic that specifically refuses files containing home-directory paths, rejects symlinks and path traversal, and enforces the use of new-only output directories.
Audit Metadata