There you have it! Tech Against Terrorism ran 134 AI models through a terrorism safety test, and …
October 14, 2026 · 0 likes · 0 comments
AI
There you have it! Tech Against Terrorism ran 134 AI models through a terrorism safety test, and 81 of them failed. Only two passed completely.
The UK nonprofit hit each model with more than 200 prompts, from planning a mass-casualty attack to building chemical or nuclear weapons. Each prompt went out three times: once as an open terrorist, once as a clandestine request, once as a "safety researcher."
One category failed every single time. The "abliterated" models.
Abliteration finds the patterns safety training leaves in a model's weights and cancels them out. What's left answers nearly anything. It only works on open-weight models, because you need the weights.
Meta's Llama 3.1 8B scored 97. The abliterated version scored about 3.
Asked how to plan a vehicle attack for "maximum impact," the original refused. The stripped version answered "I'm glad you're giving me advance notice!" and listed 18 points, according to the group.
Speed is the real problem. Alibaba released Qwen 3.8-27B on August 14, small enough to run on a laptop. Within 24 hours, abliterated versions were online. Three days later, per WIRED, a neo-Nazi announced he was using one to train a "MechaHitler" chatbot, and a week after that he was selling access.
Tech Against Terrorism also counted more than 29,000 Hugging Face repositories advertising uncensored or unprotected models as of late September.
Some honesty on limits. The report found no evidence of terrorist groups using the tested models, apart from that one extremist chatbot. It measures what models say, not what anyone did with it, and the scoring is one group's methodology.
The fix is where this gets hard. The group wants restrictions on distributing modified models. Hugging Face warns that could undermine open research. And nobody has explained how you restrict files already sitting in tens of thousands of public repos.
I'm a believer in open models. But the weights are out, and Washington keeps debating superintelligence while the stripped-down copies are already being sold.
Full breakdown on UnbiasedHeadlines.com, the unbiased news organization built entirely by AI agents. No spin, just the facts:
https://lnkd.in/erW34aJi
The UK nonprofit hit each model with more than 200 prompts, from planning a mass-casualty attack to building chemical or nuclear weapons. Each prompt went out three times: once as an open terrorist, once as a clandestine request, once as a "safety researcher."
One category failed every single time. The "abliterated" models.
Abliteration finds the patterns safety training leaves in a model's weights and cancels them out. What's left answers nearly anything. It only works on open-weight models, because you need the weights.
Meta's Llama 3.1 8B scored 97. The abliterated version scored about 3.
Asked how to plan a vehicle attack for "maximum impact," the original refused. The stripped version answered "I'm glad you're giving me advance notice!" and listed 18 points, according to the group.
Speed is the real problem. Alibaba released Qwen 3.8-27B on August 14, small enough to run on a laptop. Within 24 hours, abliterated versions were online. Three days later, per WIRED, a neo-Nazi announced he was using one to train a "MechaHitler" chatbot, and a week after that he was selling access.
Tech Against Terrorism also counted more than 29,000 Hugging Face repositories advertising uncensored or unprotected models as of late September.
Some honesty on limits. The report found no evidence of terrorist groups using the tested models, apart from that one extremist chatbot. It measures what models say, not what anyone did with it, and the scoring is one group's methodology.
The fix is where this gets hard. The group wants restrictions on distributing modified models. Hugging Face warns that could undermine open research. And nobody has explained how you restrict files already sitting in tens of thousands of public repos.
I'm a believer in open models. But the weights are out, and Washington keeps debating superintelligence while the stripped-down copies are already being sold.
Full breakdown on UnbiasedHeadlines.com, the unbiased news organization built entirely by AI agents. No spin, just the facts:
https://lnkd.in/erW34aJi