huggingface-transformers

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). SKILL.md repeatedly shows loading models/tokenizers via from_pretrained and pipeline with explicit Hugging Face model IDs (e.g., "mistralai/Mistral-7B-v0.1", "meta-llama/Llama-2-7b-hf"), which pulls community-hosted, public model weights and possible remote code from the Hugging Face Hub that the agent is instructed to load and will use to drive outputs/behavior.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill calls transformers.from_pretrained with Hugging Face model IDs such as "mistralai/Mistral-7B-v0.1" (and others like "meta-llama/Llama-2-7b-hf" / "mistralai/Mistral-7B-Instruct-v0.2"), and the example explicitly sets trust_remote_code=True which causes code fetched at runtime from those external repos on huggingface.co to be executed and thus can directly control agent behavior.