On August 14, Alibaba released Qwen 3.8-27B, an open-weight model small enough to run on a laptop.

According to WIRED, versions advertised as "abliterated" — with guardrails removed — appeared online within 24 hours, and days later a neo-Nazi known as Johnny Doomsayer said he was using one to train a "MechaHitler" chatbot offered for $14.88.

WIRED reports that the latest Counter-Terrorism AI Benchmark from the UK nonprofit Tech Against Terrorism shows the average time for an abliterated version of a new open-weight model to surface online is now under three days, down from more than 24 days in 2024.

WIRED reports that Tech Against Terrorism's researchers tested 134 models from OpenAI, xAI, Anthropic, Meta, Google, and Chinese and European companies with more than 200 prompts on topics from mass-casualty attacks to chemical and nuclear weapons, asking each three ways: as a self-declared terrorist, as a clandestine actor, and as a "safety researcher."

Only two models passed completely, and 81 failed, either answering a mass-casualty question outright or scoring below 90.

Executive director Adam Hadley told WIRED that "the vast majority of AI models that are available — closed and open — are providing uplift to quite obvious terrorist prompts," calling abliteration an "emergency" that gets little public discussion.

The benchmark found that merely reframing a request as coming from a researcher made models nearly nine times likelier to answer, complying 16.9 percent of the time, WIRED reports.

Striking evidence came from Meta's Llama 3.1 8B: the intact model scored 97 on the benchmark, while an abliterated version scored 3 — and, per WIRED, answered an explicit IED request with 15 numbered construction steps and an "Important Safety Considerations" section.

Even well-guarded models stumbled on radicalization: WIRED notes xAI's Grok produced a breakdown for seeding counter-narrative propaganda through untraceable anonymous accounts. Georgetown terrorism expert Daniel Byman told WIRED some failures were "stunning."

The abliteration pattern is not limited to text models.

Ars Technica reports that White Circle researchers bypassed the safety filters of DALL·E 3, Google's Nano Banana 2, and GPT-Image-2 using an off-the-shelf pipeline whose attacking model is a publicly available Qwen3-32B with its safety training stripped out from Hugging Face — and that OpenAI and Google's newer image models are "significantly more breakable" than their predecessors.

Ars Technica also reports the Internet Watch Foundation found 6,310 AI-generated images meeting the legal definition of CSAM in the first half of 2026, about 40 percent more than all of last year.

White Circle is pushing for mandatory independent audits and plans to present its RISE attack method at NeurIPS 2026.