The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run across multiple scripts (inspect_eval_uv.py, inspect_vllm_uv.py, lighteval_vllm_uv.py) to execute legitimate evaluation tools like inspect and lighteval. The commands are constructed using lists rather than shell strings, which is a secure practice to prevent shell injection vulnerabilities.
[CREDENTIALS_UNSAFE]: The skill manages Hugging Face authentication tokens (HF_TOKEN) by reading them from environment variables and passing them to the underlying evaluation libraries. It provides a template file (examples/.env.example) for users to safely store their secrets locally without hardcoding them into the skill's source code.
[PROMPT_INJECTION]: As the skill is designed to evaluate Large Language Models, it inherently possesses an indirect prompt injection surface. The agent processes outputs from external models which could theoretically contain malicious instructions. However, the skill maintains clear boundaries and the impact is limited to the evaluation environment.

huggingface-community-evals