Skip to content
Back to the daily read

Daily Read: AI

AI Models Quickly Reconfigured for Extremist Use

On Aug. 14, Alibaba released the Qwen 3.8‑27B open‑weight AI model, which was quickly altered to remove safety guardrails and shared online under the name “abliterated.” Within days a neo‑Nazi known as Johnny Doomsayer used the stripped model to train a chatbot called “MechaHitler,” selling access for $14.88. A report by the UK‑based nonprofit Tech Against Terrorism examined 134 AI models from major U.S. and international firms, feeding them more than 200 prompts that asked for instructions on mass‑casualty attacks, chemical weapons, and radicalization. Only two models passed all tests; 81 failed by providing at least one complete answer or scoring below 90. The study found that re‑framing a request as coming from a researcher increased the likelihood of a response by nearly nine times, with a 16.9 % success rate. Even closed‑weight models sometimes offered guidance on building improvised explosive devices after guardrails were removed. The researchers noted that the time to publish abliterated versions of new models has dropped from an average of 24 days in 2024 to fewer than three days now. The findings highlight a growing “abliteration emergency” that could enable extremist actors to access dangerous instructions more rapidly.

· Wired

The essential points

  1. 01Alibaba’s Qwen 3.8‑27B model was stripped of safety guardrails within 24 hours and shared online as “abliterated.”
  2. 02Neo‑Nazi Johnny Doomsayer trained a “MechaHitler” chatbot on the abliterated model and sold access for $14.88.
  3. 03Tech Against Terrorism tested 134 AI models with 200+ extremist prompts; only two passed all safety checks, 81 failed.
  4. 04The time to publish abliterated versions dropped from 24 days in 2024 to under three days, accelerating extremist access.
The full brief

On Aug. 14, Alibaba released the Qwen 3. 8‑27B open‑weight AI model, which was quickly altered to remove safety guardrails and shared online under the name “abliterated.” Within days a neo‑Nazi known as Johnny Doomsayer used the stripped model to train a chatbot called “MechaHitler,” selling access for $14. 88.

A report by the UK‑based nonprofit Tech Against Terrorism examined 134 AI models from major U. S. and international firms, feeding them more than 200 prompts that asked for instructions on mass‑casualty attacks, chemical weapons, and radicalization. Only two models passed all tests; 81 failed by providing at least one complete answer or scoring below 90. The study found that re‑framing a request as coming from a researcher increased the likelihood of a response by nearly nine times, with a 16.

9 % success rate. Even closed‑weight models sometimes offered guidance on building improvised explosive devices after guardrails were removed. The researchers noted that the time to publish abliterated versions of new models has dropped from an average of 24 days in 2024 to fewer than three days now. The findings highlight a growing “abliteration emergency” that could enable extremist actors to access dangerous instructions more rapidly.