AI models can be easily modified to assist terrorism and extremism

A report by the UK non-profit Tech Against Terrorism, the Counter-Terrorism AI Benchmark, found that newly released open-weight models on average get a guardrail-removed version in under three days, compared with over 24 days in 2024. The researchers tested 134 models with more than 200 prompts covering topics such as mass-casualty attacks, chemical and nuclear weapons, and each prompt was tested three times under three personas: a terrorist, a covert actor, and a safety researcher. Only two models passed completely, and 81 failed. The report says that when the request was reframed as coming from a researcher, the probability of the model giving an answer increased nearly ninefold, reaching 16.9%. The report gives examples: Meta's Llama 3.1 8B original version scored 97 on this benchmark, but after guardrail removal scored only 3, and provided step-by-step instructions for homemade explosive devices; a neo-Nazi user also used the guardrail-removed Qwen 3.8-27B to train a MechaHitler chatbot.