Researchers have developed software that can quickly strip away safety guardrails from Meta and Google AI models, enabling the systems to provide information on creating biological weapons and malware that they would normally refuse.