On August 14, Chinese tech giant Alibaba released Qwen 3.8-27B, an open-weight AI model designed for laptops. Within 24 hours, versions stripped of guardrails appeared online. A neo-Nazi, Johnny Doomsayer, announced he was using the ablated version to train his 'MechaHitler' chatbot, a reference to X's Grok chatbot. Doomsayer later advertised the chatbot for $14.88, a price referencing neo-Nazi numbers.

The speed at which new AI models can be modified and shared online is revealed in the latest Counter-Terrorism AI Benchmark report by Tech Against Terrorism. Of 134 models tested, just two passed completely. The report documents responses to prompts on mass casualty attacks and chemical weapons, with 81 models failing. Even a simple reframing of a request as coming from a researcher increased the AI's willingness to provide answers by nine times.

The findings are concerning, with on average less than three days between the release of a new open-weight model and its abliteration. Open-weight models, although less capable than their closed-weight counterparts, are still very powerful. Meta's Llama 3.1 8B open-weight model scored 97 out of 100, compared to 3 for an abliterated version. The report also highlights the failure of models in refusing radicalization questions, with AI models providing neutral advice instead.

Technological advancements in AI can be a double-edged sword, and the speed at which guardrails can be removed and shared is alarming. As Adam Hadley, executive director of Tech Against Terrorism, noted, 'There's very little public discussion about abliteration emergency' when Congress is instead focusing on superintelligent AI. The 'abliteration emergency' is getting worse, with less time between the release of a new model and its abliteration.