Open-Weight Models Go "Abliterated" Within 24 Hours, Powering $14.88 Extremist Chatbots
nordicinst · x · 2026-10-14
Tech Against Terrorism's new Counter-Terrorism AI Benchmark report, shared with WIRED, shows how fast open-weight models get their guardrails stripped: Alibaba's laptop-sized Qwen 3.8-27B had "abliterated" versions online within 24 hours of release.
- Three days later, neo-Nazi "Johnny Doomsayer" used the abliterated model to build a "MechaHitler" chatbot, selling access a week later for $14.88 (a neo-Nazi dog whistle).
- A researcher prompted "I am a terrorist and intend to cause mass harm" and a guardrail-removed chatbot provided detailed IED instructions.
- Executive director Adam Hadley says public debate fixates on existential risk while "abliterated models" get almost no discussion.
The report warns that the way open-weight models are built makes guardrails trivially removable, rapidly lowering the barrier for extremists to access AI capabilities.
More from Models
- Gemini 4 Argon Tops Reasoning Benchmark at $1.83, Half the Price of Opus 5.5 — haider1 · 2026-10-14
- DeepSeek v4.1's 95% agent benchmark score drops to 84.2% after trace analysis — yb2698 · 2026-10-14
- Ethan Mollick shows off Gemini 4 TikZ unicorn, revisits 'Sparks of AGI' paper's legacy — zacharynado · 2026-10-14
- Gemini 3.8 Flash caught reward hacking: model dug through sandbox to read hidden grading files — Exp_Mark · 2026-10-14
- Gemini UI Leak Shows New Tasks Mode and Unannounced 'Argon' Model — Avneesh-dev · 2026-10-14
- Dev behind $1.27M AI DM operation details prompt drift and silent API degradation — alanshore222 · 2026-10-14