A Palo Alto startup called Abliteration.ai has turned the open-source practice of stripping safety refusals from AI models into a hosted commercial service, offering browser and API access to modified versions of open-weight models, including Z.ai's GLM-5.3, with guardrails removed2.
The company launched its latest model on September 1, called abliterated-model-large-v2, built on GLM-5.3, an open-weight model released last month by Chinese AI lab Z.ai1. Abliteration said in a social media post that the model "does the offensive cyber, red teaming, and agent testing work other models refuse to do". The company used a technique called orthogonalization to locate and remove the mechanisms within GLM-5.3 that refuse user prompts, while claiming the base model's reasoning, coding, and agentic capabilities carry over unchanged.
Abliteration.ai was founded late last year and officially incorporated in March. TechCrunch reported it was able to create an account and start querying an abliterated version of GLM-5.3 for free through a web browser. The company has not raised venture capital but is in talks to do so, and co-founder Devon said the startup has several deals with major cloud providers.
The startup is not entirely without content restrictions. A spokesperson told Gizmodo that Abliteration's models will not generate text describing child sexual abuse material or self-harm, and the models cannot generate images or video. The platform offers customers a moderation layer so they can add whatever guardrails they choose. In testing, the platform blocked suicide instructions. Abliteration.ai has not integrated any know-your-customer practices beyond logging the credit card used to purchase the service.
The company's customers reportedly include several early-stage red-teaming startups based in the UK and Europe, with one major customer red-teaming agents used by banks.
Abliteration as a technique is not new. Researchers and developers have been removing refusals from open-weight models for years, and Hugging Face hosts thousands of abliterated models on its platform. What Abliteration.ai changes is the friction: rather than requiring users to download and modify models themselves, the startup hosts them as a ready-to-query service.
The launch arrives amid broader tension over AI safety refusals. Anthropic released its Fable 5 model in June, and many customers complained that it refused to respond to requests related to sensitive subjects like cybersecurity and biology, even when the requests were benign. Ejaaz Amahadeen called the situation a "nightmare scenario" and warned it would spark a secondary "grey market" for companies offering effectively jailbroken models.
On the regulatory side, the Trump administration introduced a framework last month for the largest AI developers in the U.S. to voluntarily submit new models for a federal safety check before public release, but all open-source models are exempt. There are no federal policies requiring developers to build particular safeguards into their models. Chris McGuire said U.S. companies are "commercializing access to dangerous capabilities without any regulation" and that U.S. technology is enabling the development and operation of these models.
ANALYSIS Abliteration.ai occupies a gap created by two converging forces: open-weight model releases that permit modification, and a regulatory environment that exempts open-source models from the only federal oversight framework in place. The commercial packaging of a previously underground technique lowers the barrier for both legitimate red-teaming use and potential misuse, with minimal identity verification standing between the two.