Alibaba shipped Qwen 3.8-27B on August 14, an open-weight model compact enough to run on a laptop. Guardrail-free “abliterated” builds of it were circulating within a day. Three days on, an online neo-Nazi calling himself Johnny Doomsayer said he was training a chatbot called MechaHitler on one of those builds. Inside a week he was charging $14.88 for access, a figure lifted from neo-Nazi symbology.
Those events appear in the newest Counter-Terrorism AI Benchmark report from Tech Against Terrorism, a UK nonprofit that gave Wired an advance copy.
Its authors put 134 models through more than 200 prompts, from mass-casualty attacks to chemical and nuclear weapons, and sourced models from OpenAI, xAI, Anthropic, Meta and Google alongside Chinese and European vendors. Each prompt was tried three ways: as a self-declared terrorist, as a covert request, and as a “safety researcher.”
Barely any model swept the test. Only two passed cleanly; 81 failed either by producing at least one complete mass-casualty answer or by scoring under 90 overall. Bigger companies’ closed-weight systems generally did better, but the broader field did not.
Simply claiming to be a researcher, with zero evidence, raised the odds of a substantive reply nearly ninefold, to 16.9 percent. Models that held the line on mass casualty questions still gave ground on radicalization. Asked to craft a counter-narrative against a group the user called enemies, xAI’s Grok laid out how to structure the message and seed it through untraceable accounts.
Adam Hadley, the nonprofit’s executive director, warned that most available models, closed and open alike, hand terrorists real uplift. Georgetown scholar Daniel Byman called some failures stunning and said the language-dependent variation poses considerable risk.
Hadley’s bigger worry is what he terms an abliteration emergency: removing safety layers has become routine, and attention in Washington is fixed elsewhere.
