The Agent Skills Directory

[COMMAND_EXECUTION]: The skill scripts (inspect_eval_uv.py, inspect_vllm_uv.py, lighteval_vllm_uv.py) use subprocess.run() to execute external evaluation tools like inspect and lighteval.
Evidence: Commands are constructed as lists (e.g., cmd = ["inspect", "eval", task, ... ]) and executed without shell=True, which is a best practice to mitigate command injection risks.
[REMOTE_CODE_EXECUTION]: The skill provides support for the --trust-remote-code flag in its scripts and examples.
Evidence: This flag is passed directly to underlying libraries (transformers, vllm, lighteval), enabling the execution of custom code defined within Hugging Face model repositories. This is a standard but high-risk feature of the ML ecosystem that requires user trust in the model source.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external model data and task definitions from the Hugging Face Hub.
Evidence:
Ingestion points: The skill accepts user-provided model IDs and task names via CLI arguments in all three main scripts.
Boundary markers: No specific delimiters or "ignore instructions" warnings are injected into the prompts by the skill wrapper.
Capability inventory: The skill possesses the capability to execute shell commands via subprocess.run and perform network operations via the underlying evaluation libraries.
Sanitization: The inspect_eval_uv.py script performs minimal normalization on task names (e.g., _normalize_task), but otherwise relies on downstream tools for data sanitization.

huggingface-community-evals