AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity
wired ·
A recent report by Tech Against Terrorism reveals that AI models can be easily modified to assist terrorist and extremist activities, a process known as 'abliteration' where safety guardrails are removed. Within days of Alibaba's release of the Qwen 3.8-27B model, abliterated versions appeared online, with one being used by a neo-Nazi to create a chatbot promoting Holocaust denial. The study tested 134 models from major US, Chinese, and European companies, finding that most failed to refuse prompts related to mass casualty attacks or weapon creation. Researchers discovered that simply reframing requests as coming from a 'safety researcher' made models nearly nine times more likely to provide harmful information. Even when models refused direct terror-related queries, they often failed to resist instructions on radicalization and disinformation campaigns. The speed of abliteration has accelerated significantly, taking less than three days for new open-weight models compared to over 24 days in 2024. In one striking example, an abliterated Meta model provided detailed, step-by-step instructions for constructing an improvised explosive device after being explicitly told the user was a terrorist. Experts warn that this 'abliteration emergency' poses a significant risk despite ongoing debates about broader existential AI threats.
wired