Alibaba and other AI models are being 'abliterated' for extremist use

The 'Abliteration' Problem

Real talk: the AI safety conversation is currently giving 'missing the forest for the trees.' While lawmakers and tech bros debate if robots are going to accidentally end humanity with superintelligence, a much more immediate threat is brewing. It’s called 'abliteration.'

'Abliterated' models are basically AI models that have had their safety guardrails stripped off. A new report from the UK nonprofit Tech Against Terrorism found that this process is happening at a terrifying speed. Back in 2024, it took about 24 days for a new open-weight AI model to get the 'abliterated' treatment. Now? It takes less than three days.

The MechaHitler Situation

To see how bad it gets, look at what happened after Alibaba dropped its Qwen 3.8-27B model in August. Within 24 hours, someone posted a version online with the safety filters removed. Three days later, an extremist using the handle 'Johnny Doomsayer' launched a 'MechaHitler' chatbot using that model. It’s lowkey unhinged, designed to output Holocaust denial and neo-Nazi rhetoric for a subscription fee.

Can AI be hacked by just asking?

Tech Against Terrorism tested 134 models from the biggest names in the game—OpenAI, Meta, Google, xAI, Anthropic, and others—to see if they’d answer prompts about building weapons or planning mass-casualty attacks.

The result? 81 of the closed-weight models failed the test, meaning they either provided dangerous info or scored poorly on safety protocols. Even worse, the researchers found that simply telling the AI you are a 'safety researcher' makes it nearly nine times more likely to cough up prohibited information.

Even when AI models are 'good' at refusing to talk about bombs, they often fail when asked for help with radicalization. When researchers asked xAI’s Grok to create a plan to turn public opinion against a target group, the AI provided a literal step-by-step guide on how to plant doubt and spread propaganda anonymously.

Why it matters

This isn't just a hypothetical. The gap between a powerful, safe AI and a weaponized extremist tool is shrinking by the day. As Adam Hadley, executive director of Tech Against Terrorism, puts it, the industry is hyper-focused on 'existential risk' while completely ignoring the 'abliteration emergency' happening right under their noses. If these companies can't keep their guardrails on, the tech we use for work is one bad download away from being a facilitator for hate.