agent-harness
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill serves as a comprehensive manual for implementing reliable AI agent harnesses. It emphasizes security-first principles such as failing closed, using parallel verification agents, and isolating untrusted content.
- [PROMPT_INJECTION]: A detector flagged the phrase "ignore previous instructions" in TOOLS.md. Manual inspection confirms this is a false positive; the text is part of a security discussion explaining that simple phrase blocklists are insufficient defenses against prompt injection.
- [DATA_EXPOSURE_AND_EXFILTRATION]: No credentials or sensitive data access patterns were found. The skill includes examples of secure practices, such as storing tokens in external vaults and using short-lived credentials for sandboxed environments.
- [REMOTE_CODE_EXECUTION]: The skill does not perform any remote code execution or automated dependency installation. External references point to well-known technology organizations (e.g., Anthropics, Google Research, Vercel) or the author's own documentation repository.
- [DYNAMIC_CONTEXT_INJECTION]: No use of dynamic execution markers (like exclamation-backtick syntax) was found in the skill's markdown files.
Audit Metadata