prompt-guard

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

• Third-party content exposure detected (high risk: 0.85). Workflow 2/3 explicitly “filter third-party data (API responses, web scraping, RAG docs)” by passing that outsider-authored free text into the classifier (filter_third_party_data(data, ...) / batch_filter_documents(documents, ...) → tokenizer(text, ...) → model logits), which is LLM-readable prose at runtime.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill calls AutoModelForSequenceClassification.from_pretrained with the Hugging Face model meta-llama/Prompt-Guard-86M (https://huggingface.co/meta-llama/Prompt-Guard-86M), which downloads remote model artifacts at runtime that directly determine prompt-filtering decisions and are required for the skill to operate.