cuda-kernels

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • External Binary Downloads: The skill facilitates the use of the get_kernel utility to download and load pre-compiled CUDA kernels from the HuggingFace Hub. This functionality is intended to streamline the deployment of optimized kernels across different environments without requiring local C++ compilation. Users should ensure they are loading kernels from trusted repositories.
  • Command Execution via Build Tools: The instructions provide examples for using build systems like Nix and the kernel-builder tool to compile CUDA source code into Python extensions. These operations involve executing shell commands to manage the build lifecycle, which is an expected part of a developer-focused workflow for GPU optimization.
  • Indirect Prompt Injection Surface: Several benchmarking and integration scripts are designed to process user-supplied text prompts for generating video or text. While these scripts use established library pipelines (such as diffusers and transformers), the ingestion of external data represents a standard interface that could be affected by specially crafted inputs.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 05:29 PM
Security Audit — agent-trust-hub — cuda-kernels