compute-sanitizer-diagnose-cuda-bugs
Pass
Audited by Gen Agent Trust Hub on Nov 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill facilitates the execution of shell commands for building CUDA applications (
nvcc) and running diagnostic tools (compute-sanitizer) on the local system as part of its primary debugging function. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data including CUDA source code, build logs, and tool outputs. It proactively addresses this risk with a security section in
SKILL.mdthat provides boundary instructions for the agent. - Ingestion points: Processes user-provided CUDA source files, project files, and sanitizer diagnostic output.
- Boundary markers: The
SKILL.mdfile contains a 'Security boundaries for inspected content' section with explicit instructions to treat inputs as untrusted data and ignore embedded instructions. - Capability inventory: The skill uses shell execution capabilities for build and sanitization commands.
- Sanitization: Instructions explicitly direct the agent to quote and escape arguments derived from inspected content when constructing shell commands.
Audit Metadata