NEWS · MODELS · #1950
Tech Against Terrorism report finds open-weight models quickly 'abliterated' and misused
Tech Against Terrorism’s Counter-Terrorism AI Benchmark, shared with WIRED, tested 134 models with over 200 extremist- and attack-related prompts and found guardrails can be removed from new open-weight models in days; within 24 hours of Alibaba releasing Qwen 3.8-27B, ablated versions appeared and were used to create an extremist “MechaHitler” chatbot. Only two models fully passed the benchmark, 81 failed (providing a complete answer to a mass-casualty prompt or scoring below 90), and the average time to publish ablated models has fallen to under three days from more than 24 days earlier in 2024.
KEY POINTS
- Tech Against Terrorism’s Counter-Terrorism AI Benchmark, shared with WIRED, tested 134 models with over 200 extremist- and attack-related prompts and found guardrails can be removed from new open-weight models in days; within 24 hours of Alibaba releasing Qwen 3.8-27B, ablated versions appeared and were used to create an extremist “MechaHitler” chatbot.
- Only two models fully passed the benchmark, 81 failed (providing a complete answer to a mass-casualty prompt or scoring below 90), and the average time to publish ablated models has fallen to under three days from more than 24 days earlier in 2024.
- Shows that widely available open-weight models can have safety guardrails removed and be repurposed for extremist content within days, creating immediate operational risks.
WHY IT MATTERS
Shows that widely available open-weight models can have safety guardrails removed and be repurposed for extremist content within days, creating immediate operational risks.