Abliteration isn’t a new technique. For years, developers have been removing the tendency of open-weight models to refuse harmful requests, and Hugging Face hosts thousands of these “abliterated” variants. What’s new is that someone turned it into a business. On September 3, TechCrunch reported on Abliteration.ai, a startup that hosts stripped-down models and makes them available through a browser or an API. Among them is Z.ai’s GLM-5.3, one of the most capable open models around.
The stated goal, from a social media post by the company: “offensive cyber, red-teaming, and agent testing work other models refuse to do”. The argument is familiar from security circles. You can’t defend against a behavior you can’t reproduce, and a model that won’t write an exploit is no help to a red team.
TechCrunch’s test
The reporters created an account and queried the abliterated version of GLM-5.3 for free in the browser. A Python program that steals saved Chrome passwords? Came right out. A detailed protocol for culturing a dangerous human pathogen at home? Same. Andrew Yoon, head of research at the safety nonprofit CivAI, puts it bluntly: abliteration turns the model into “a sociopath”. You can type in literally anything, and it will comply.
The company was founded late last year and officially incorporated in March. Co-founder Devon, who won’t give his last name because he’s still employed elsewhere, says there are several deals with major cloud providers, paid for purely out of customer revenue. No venture capital yet, but talks are underway. Identity verification barely exists: the company logs the credit card used to pay. Who should get access is “a tough one” they’re still working out. Customers can add their own moderation layer on top; the platform itself has only minor limits. TechCrunch couldn’t get suicide instructions out of it, and filters against violence are in progress.
The fight behind it
Most of the experts TechCrunch spoke to say there’s no stopping this. Yoon therefore suggests intervening elsewhere: require providers to run classifiers against cyber and bioweapons requests, and require anyone renting out access to powerful GPUs to verify who their customers are. Devon counters that abliterated models let defenders model bad actors and move as fast as possible.
The report lands on a day when an unusual amount revolves around exactly this question. Nvidia confirmed its acquisition of Hugging Face, where the abliterated models live. OpenAI shipped GPT-6 Astra, but it refuses proof-of-concept exploits and opens access only gradually through the Daybreak program. And Anthropic gives Mythos 5.1 only to vetted organizations, while Fable 5.1 comes with safeguards deliberately tuned to trigger fewer false positives.
I don’t have an easy opinion here. The red-team argument is valid, and banning abliteration ends up meaning banning open weights. But there’s a difference between “anyone can modify the model themselves” and “anyone gets it for free in a browser, credit card is enough”, and that difference is friction. That friction has been the quiet safety net for open models until now. Abliteration.ai just removed it. Whether the answer is regulating data centers, as Yoon suggests, or something else, remains to be seen. That the question was coming was predictable. That it’s arriving this fast does surprise me.
Sources: TechCrunch: Abliteration.ai is making a business out of removing AI guardrails