The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/benchmark_via_builtin.py uses subprocess.Popen to execute flashinfer_benchmark.py within a directory specified by the user via the --flashinfer_repo argument. This allows the execution of arbitrary code contained within the targeted directory if a user provides an untrusted path.
[REMOTE_CODE_EXECUTION]: In scripts/benchmark_model.py, the _load_meta_model function exposes the --trust-remote-code parameter of the Hugging Face transformers library. When enabled, this feature allows the execution of arbitrary Python code defined in the model repository. The SKILL.md file contains a mitigation instruction: 'Do not enable --trust-remote-code without explicit approval; pin --revision <commit> when it is needed.'
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted model configuration data from the Hugging Face Hub, creating an attack surface where maliciously crafted model metadata could influence agent behavior.
Ingestion points: scripts/benchmark_model.py via AutoConfig.from_pretrained.
Boundary markers: Absent in prompt interpolation for derived shapes.
Capability inventory: Execution of shell commands via subprocess, file system writes to a work directory, and network access to Hugging Face.
Sanitization: The skill uses shlex.join to safely construct command strings for display and execution.

modelopt-benchmark-model-kernels