rocm-kernels
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- External Resource Ingestion: The skill utilizes the kernels library and Hugging Face Hub to fetch pre-compiled binaries and model weights. These resources originate from the vendor's own verified ecosystem, which is a standard source for these optimization assets.
- Indirect Prompt Injection Surface: The benchmarking utility in scripts/benchmark_e2e.py accepts user-supplied prompts via command-line arguments.
- Ingestion points: User input enters the system through the prompt parameter in the benchmark runner within scripts/benchmark_e2e.py.
- Boundary markers: The implementation does not currently utilize specific boundary delimiters or safety instructions when interpolating user-provided text into the generation pipeline.
- Capability inventory: The skill possesses capabilities to write to the file system, including saving JSON results and exporting MP4 video files, alongside performing GPU-accelerated computations.
- Sanitization: The skill does not apply specific validation or sanitization to prompt strings before they are processed by the machine learning models.
- Dynamic Execution: The skill facilitates the runtime compilation of Triton kernels and the loading of external binary modules. This is an essential function for GPU kernel development and is performed using recognized machine learning libraries.
Audit Metadata