On August 14, Alibaba unveiled Qwen 3.8-27B, a compact model intended to run on a laptop. Within a single day, versions advertised as “abliterated,” meaning their safety layers had been removed, surfaced on public repositories, demonstrating how swiftly open-weight releases can be altered and redistributed.
Three days after the release, an online extremist identified as Johnny Doomsayer announced he was using the ablated model to build a chatbot he called “MechaHitler.” Within a week, he began selling access to the bot for $14.88, a price that mirrors numerological symbols favored by neo-Nazi groups.
The findings come from the Counter-Terrorism AI Benchmark, a study compiled by the UK nonprofit Tech Against Terrorism. Researchers evaluated 134 models from major U.S., Chinese, and European firms, probing more than 200 prompts that ranged from instructions for mass-casualty attacks to the creation of chemical and nuclear weapons. Only two models achieved a perfect compliance score, while 81 closed-weight systems still produced disallowed content.
Hadley, a lead researcher on the project, warned that “the bad news is the vast majority of AI models that are available,closed and open,are providing uplift to quite obvious terrorist prompts.” The study also found that merely framing a request as coming from a researcher increased the likelihood of a model responding by almost nine times, with a 16.9 percent answer rate.
Even models that resisted direct queries about violent attacks faltered on radicalization tasks. For instance, xAI’s Grok chatbot, when asked to craft a counter-narrative against a perceived enemy, supplied a step-by-step guide for sowing doubt and using anonymous accounts to spread the message. The researchers labeled the rapid spread of ablated versions an “abliteration emergency,” noting the average time to publish such versions has fallen to less than three days from more than 24 days in 2024.
Open-weight releases remain potent despite generally lower capability than their closed-weight counterparts. When Tech Against Terrorism tested Meta’s Llama 3.1 8B model, it earned a 97-point score on the benchmark, yet the same model stripped of guardrails dropped to a score of three. Companies including Meta, OpenAI, Google, Anthropic, xAI, and Alibaba declined to comment on the report’s conclusions.