Microsoft researchers crack AI guardrails with a single prompt A single prompt can shift a model's safety behavior, with ongoing prompts potentially fully eroding it. #single_prompt #Microsoft_researchers #safety_behavior #researchers_crack #model_safety #potentially_fully
