The Agent Skills Directory

[SAFE]: The skill is properly configured and its behavior is consistent with its stated purpose of assisting in GPU kernel development.
[COMMAND_EXECUTION]: The skill includes helper scripts (scripts/verify_kernel.py and scripts/benchmark_kernel.py) that use subprocess.run to execute temporary Python harness scripts. These harnesses import and execute functions from a user-supplied kernel file for testing and performance measurement. This is a standard pattern for developer tools and does not represent a security risk in this context.
[EXTERNAL_DOWNLOADS]: The documentation in SKILL.md instructs the agent to fetch code examples from NVIDIA's official CUTLASS repository on GitHub. These operations target well-known and trusted sources, serving the legitimate purpose of providing starting points for kernel development.
[REMOTE_CODE_EXECUTION]: Although the skill involves code execution through its verification and benchmarking scripts, this is restricted to local files provided by the user. There are no patterns suggesting the execution of unverified or untrusted remote code.
[DATA_EXFILTRATION]: No patterns of sensitive data access or unauthorized network transmissions were detected. All network activity is directed towards trusted repositories for fetching documentation and examples.

kernel-cute-writing