llamaguard

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

• Potentially malicious external URL detected (high risk: 0.80). The skill loads the external model "meta-llama/LlamaGuard-7b" at runtime via calls like AutoModel.from_pretrained and vLLM (which fetch from https://huggingface.co/meta-llama/LlamaGuard-7b), so required remote model weights—fetched during runtime—directly determine the agent's moderation outputs.