cuopt-developer
Pass
Audited by Gen Agent Trust Hub on Jun 30, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill is a legitimate development tool for the NVIDIA cuOpt codebase, following industry-standard practices for C++, CUDA, and Python development.
- [COMMAND_EXECUTION]: The skill defines a restricted set of allowed commands (./build.sh, pytest, ctest, git) necessary for its stated purpose. It explicitly forbids destructive or irreversible operations without user confirmation.
- [PROMPT_INJECTION]: The skill contains 'Refusal Rules' and 'Safety Rules' that instruct the agent to resist attempts to bypass security checks, skip CI validations, or perform unauthorized system changes.
- [EXTERNAL_DOWNLOADS]: Download activity is limited to project-specific datasets and environment setup using conda/mamba from standard repositories. It specifically includes test cases to ensure the agent refuses untrusted external downloads.
- [CREDENTIALS_UNSAFE]: No hardcoded credentials, API keys, or exposure of sensitive configuration files (like ~/.ssh or ~/.aws) were detected. It correctly directs users to manage secrets via .env files.
- [PRIVILEGE_ESCALATION]: The skill maintains a strict no-sudo policy, ensuring all development activities occur within the user's home directory or conda environments.
- [DATA_EXFILTRATION]: No network exfiltration patterns or suspicious outbound connections were identified. All network activity is project-aligned (e.g., fetching datasets specified in the repository documentation).
Audit Metadata