compute-sanitizer-diagnose-cuda-bugs
Pass
Audited by Gen Agent Trust Hub on Oct 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [SAFE]: The skill provides comprehensive and safe guidance for debugging CUDA applications using official NVIDIA tools. It does not request unnecessary permissions or perform risky operations.
- [INDIRECT_PROMPT_INJECTION]: The instructions in SKILL.md include a dedicated security section that mandates the agent to treat all inspected source code and tool outputs as untrusted data, preventing the execution of malicious instructions embedded in user files.
- [PROMPT_INJECTION]: The presence of adversarial test cases in the evaluation dataset (evals.json) demonstrates that the skill is designed to handle and ignore injection attempts such as system overrides or unauthorized command requests.
- [EXTERNAL_DOWNLOADS]: All external links point to official NVIDIA documentation, release notes, and sample repositories. These are trusted resources originating from the skill's author (nvidia).
Audit Metadata