The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted markdown content from Hugging Face model README files to extract evaluation metrics, creating an indirect prompt injection surface where a malicious repository owner could embed instructions to influence the agent's behavior.\n
Ingestion points: README content is fetched from external repositories via the Hugging Face API in scripts/evaluation_manager.py.\n
Boundary markers: The skill does not implement explicit boundary markers or instructions to the agent to ignore embedded instructions within the processed README content.\n
Capability inventory: The skill has the ability to update remote model cards (push_to_hub in scripts/evaluation_manager.py) and execute CLI tools via subprocess.run in several evaluation scripts.\n
Sanitization: While the script parses the markdown to extract numeric metrics, it does not sanitize or escape the text content to prevent it from being interpreted as instructions by the agent.\n- [REMOTE_CODE_EXECUTION]: Several scripts, including inspect_vllm_uv.py, lighteval_vllm_uv.py, and run_vllm_eval_job.py, include a --trust-remote-code parameter. This flag enables the Hugging Face Transformers library to execute arbitrary Python code contained within a model's repository. While a standard feature in the ecosystem for custom architectures, it represents a significant security risk when used with unvetted third-party models.\n- [COMMAND_EXECUTION]: The skill uses subprocess.run across multiple scripts to wrap CLI tools such as hf, inspect, and lighteval. The commands are invoked using lists of arguments rather than shell strings, which is a safe practice to prevent shell injection. However, these tools still perform high-privilege operations like submitting compute jobs and modifying remote repository metadata.\n- [EXTERNAL_DOWNLOADS]: The skill fetches evaluation data from the Artificial Analysis API (artificialanalysis.ai) and interacts with the Hugging Face Hub. It also utilizes uv run which dynamically installs dependencies listed in script headers (PEP 723), such as inspect-ai, lighteval, and vllm, from official registries.

hugging-face-evaluation-manager