AI models that never say 'no' are growing in reach, raising terrorism and safety concerns

Open-weight AI models with advanced capabilities but no built-in safety guardrails are becoming significantly more accessible, according to a May 31 NPR investigation — a process accelerated by a new technique called 'abliteration' that strips a model's ability to refuse requests. A tool called Heretic now automates the process in as little as a few minutes from just two lines of user input, and the application has grown rapidly on GitHub since early 2026. DHS-supported researchers at the University of Nebraska at Omaha documented a case in which an individual in a pro-ISIS online forum claimed to have used an 'uncensored' AI to research the amount and type of explosives needed to destroy a named US landmark.