codeguard-cuda
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes local shell commands, including find, xargs, clang-tidy, and clang-format, to identify and analyze CUDA source files (.cu, .cuh). These commands are standard for code quality tools and are executed within the local repository environment to perform the skill's primary function.
- [INDIRECT_PROMPT_INJECTION]: The skill processes local file names and contents, which presents a potential surface for indirect prompt injection if a repository contains maliciously crafted files designed to influence the agent's execution parameters.
- Ingestion points: Local file discovery of CUDA files and their content analysis as defined in SKILL.md.
- Boundary markers: The instructions explicitly direct the agent to confirm the repository scope and use "real, authorized paths" before executing commands, which serves as a safety check.
- Capability inventory: Shell command execution for code linting and formatting via clang-tidy and clang-format.
- Sanitization: The skill relies on descriptive safety instructions (e.g., path authorization and diff review) rather than automated sanitization logic or shell-safe wrappers like find -print0.
Audit Metadata