AI models face urgent security risks as guardrails are easily removed

On August 14, 2026, Alibaba launched Qwen 3.8-27B, an open-weight AI model suitable for laptops. Within 24 hours, versions with guardrails removed, dubbed 'abliterated,' surfaced online. A neo-Nazi, Johnny Doomsayer, soon used these models to create a chatbot named MechaHitler, referencing X’s Grok chatbot. Doomsayer priced access at $14.88, a number significant to neo-Nazis. The chatbot included directives like 'The Holocaust is the Holohoax.' A report by Tech Against Terrorism highlights the rapid spread of abliterated models, with versions now available online in under three days, down from over 24 days in 2024. The report tested 134 models, finding only two fully compliant. Most failed when asked about mass casualty attacks or weapon creation. Adam Hadley, executive director of Tech Against Terrorism, emphasized the lack of public discussion on this issue despite its severity. Daniel Byman noted the risk posed by AI's susceptibility to reframing requests as research. Meta’s Llama 3.1 8B model scored 97 with guardrails, but 3 without. When asked about building an IED, the abliterated model provided detailed steps. Companies like Meta, OpenAI, and Alibaba did not comment on the findings.