The Agent Skills Directory

[Dynamic Code Execution]: The skill uses importlib.util in scripts/benchmark.py and scripts/xpu_profiler.py to dynamically load and execute Python code from user-supplied Triton kernel files. While this is the intended purpose of a benchmarking tool, it allows the execution of arbitrary code within the agent's environment.
[Command Execution]: The scripts/xpu_profiler.py script executes shell commands via subprocess.run to interface with the Intel VTune Profiler. The script passes arguments derived from configuration and generated runner scripts to the subprocess.
[Privilege Escalation Considerations]: The skill's documentation (e.g., in README.md and references/workflow_details.md) suggests that users may need to modify system-level GPU observation settings using sudo to enable certain hardware profiling features. This is a common requirement for low-level performance analysis but involves sensitive system changes.
[External Dependencies]: The skill specifies a dependency in scripts/requirements.txt on the ai-bench harness hosted on GitHub. This is a well-known benchmarking tool relevant to high-performance computing, and its use here is consistent with the skill's primary purpose.
[Indirect Prompt Injection Surface]: The skill includes an optimization workflow that ingests and analyzes external PyTorch and Triton code. This creates a surface for indirect prompt injection, although the scripts primarily use static analysis (AST parsing) and controlled benchmarking to process the data.

xpu-kernels