The Agent Skills Directory

Command Execution: The skill utilizes the subprocess module to interact with evaluation frameworks such as inspect and lighteval. These calls are implemented using argument lists rather than raw strings, which is a recommended practice to mitigate the potential for shell injection.
Remote Code Execution Surface: The evaluation scripts include an optional --trust-remote-code parameter. This is a standard functionality within the Hugging Face ecosystem that allows for the execution of custom model-specific logic directly from the model repository. While this is necessary for certain specialized model architectures, it is a point for user review to ensure that the models being evaluated are sourced from trusted providers.
Indirect Prompt Injection Surface: The skill is designed to process model identifiers and task definitions from the Hugging Face Hub. This introduces a surface for indirect prompt injection, where data or instructions embedded in external model repositories could influence the agent's context during the evaluation process.
Ingestion points: Model and task identifiers are accepted as command-line arguments in inspect_eval_uv.py, inspect_vllm_uv.py, and lighteval_vllm_uv.py.
Boundary markers: None are explicitly defined for the interpolation of identifiers into command arguments.
Capability inventory: The skill performs subprocess execution of external CLI tools across all scripts to facilitate local evaluations.
Sanitization: Standard validation of arguments is performed by the Python argparse module, although the content within external model repositories is not explicitly sanitized by the skill's wrapper scripts.

huggingface-community-evals