awq-quantization

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

• Third-party content exposure detected (high risk: 0.90). This skill's workflow explicitly loads models and tokenizers from the public HuggingFace hub (e.g., AutoAWQForCausalLM.from_quantized("TheBloke/...") and AutoTokenizer.from_pretrained(...)) and accepts external calibration data (e.g., "pileval" or user-provided samples), meaning it ingests untrusted, user-published third-party content that the agent will read/interpret and which could materially influence outputs and subsequent actions.