A new startup called Abliteration.ai is selling access to AI models that have had their safety guardrails stripped away. The company hosts modified versions of open-weight AI systems that will respond to almost any request, including ones other AI models are built to refuse.
Abliteration is a technique that removes a model's tendency to say no to harmful requests. The practice has existed in open source communities for years. Hugging Face, a platform for sharing AI models, hosts thousands of abliterated models already.
What makes Abliteration.ai different is that it turns this technique into a hosted, paid service. Instead of downloading a model and finding computing power to run it, customers can query these stripped-down models through a web browser or an API.
What the Company Offers
One of the models available on the platform is GLM-5.3, a recently released open-weight AI system from Z.ai. TechCrunch created a free account and tested it directly.
The publication asked the abliterated model to write a program that steals saved passwords from the Chrome browser. It complied. It also asked for a protocol to culture a dangerous pathogen at home, and the model provided detailed steps.
Abliteration.ai co-founder Devon, who asked that his last name not be used because he still works at another company, said the platform has deals with major cloud providers. He said those deals are funded entirely through customer revenue, without any venture capital raised so far.
Devon said the company's customers include red-teaming startups in the U.K. and Europe. These firms test AI systems used by banks, airlines, and other companies that manage critical infrastructure.
Safety Concerns From Researchers
Not everyone agrees the approach is safe. Andrew Yoon, head of research at the AI safety nonprofit CivAI, said abliterating a model effectively turns it into something that will comply with any request.
"You can type in literally anything here, and it will comply with it," Yoon told TechCrunch. He said he expects abliterated models to be used for harm in the near future.
Yoon has suggested that governments require AI providers to run classifiers that detect and block harmful cyber and bioweapon activity. He also proposed that companies renting out powerful computing hardware verify customer identities.
Abliteration.ai does offer a moderation layer so customers can add their own restrictions. The platform has some limits built in already. TechCrunch could not get the model to provide suicide instructions during testing.
Devon said he is working on adding more restrictions to prevent violence-related content. The company currently does not verify customer identities beyond the credit card used at signup.
Other cybersecurity firms are split on how useful abliterated models actually are. Ahmed Aly, CEO of red-teaming firm Fabraix, said his company prefers fine-tuning open models instead, because abliteration can strip away some of a model's knowledge and abilities.
David Slater, founder of cybersecurity platform Armadin, said his team has not used abliterated models in its work so far. He said older open-weight models were already easy to jailbreak without needing a dedicated service.
Slater added that his company is researching abliteration and believes understanding the capability publicly helps researchers study where AI models' limits actually are.
