eval-author-environment
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill is designed around security-first principles for agent evaluation. It includes explicit instructions to enforce network isolation (e.g., using 'no-network' modes for verifiers), avoid reading sensitive '.env' files, and reference credentials only by variable names to prevent secret exposure.
- [SAFE]: The included script 'scripts/trace_tools.py' is a legitimate entry point for the NVIDIA 'eval-author' ecosystem, providing a wrapper for environment management and task validation tools. It does not contain any malicious patterns such as obfuscation, unauthorized network calls, or persistence mechanisms.
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a potential attack surface because it processes external repository code, Dockerfiles, and trace data (ATIF) to build environments. However, it incorporates robust mitigation strategies, including the use of separate, offline verifier containers, strict data boundary markers, and independent reference queries to verify agent actions without relying on the agent's own reports. These measures effectively prevent processed data from influencing the agent's core instructions in a way that would bypass safety protocols.
Audit Metadata