quantizing-models-bitsandbytes

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

• Third-party content exposure detected (high risk: 0.90). The skill's workflows explicitly load public models and user-generated datasets (e.g., AutoModelForCausalLM.from_pretrained("meta-llama/...") and load_dataset("timdettmers/openassistant-guanaco") from the HuggingFace/model & dataset hubs), which are open third-party sources used as training/data inputs and therefore could contain untrusted instructions that materially influence subsequent model behavior.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill's training example calls load_dataset("timdettmers/openassistant-guanaco"), which at runtime fetches the external Hugging Face dataset (https://huggingface.co/datasets/timdettmers/openassistant-guanaco) containing instruction prompts that would directly control the model's fine-tuning/prompting.