Abliteration AI sells open-weight models with refusals removed

Wired's profile starts with Devon Thomas, a 36-year-old engineer who grew up in Indiana, studied computer science at Indiana University, Bloomington, and joined Apple in 2020. Members of his bible study group would call him unfailingly kind and polite. Yet his company, Abliteration AI, removes safeguards from open-weight models and makes them available to anyone. The author, who covers AI and constantly hears about the need for more safety, says the reporting challenged his assumption that uncensored models are mainly the domain of criminals.

Thomas's rebellion began in November 2025. His day job at Apple involved working with AI, but outside work he disliked the models hosted by the big labs: using them meant potentially sharing information with the labs, and the models could change from one day to the next. Worse, they refused useful work while he prototyped ideas, rejecting inputs that contained any user information. His requests were blocked and his account banned when he tried to use a model to analyse logs of cyber attacks, or to work out which users of an app might be most likely to pay for a new feature. ChatGPT and Claude steer clear of a long list of subjects, including cybersecurity requests the models judge could be malicious, anything deemed too adult, and politically sensitive topics, and open-weight models carry similar guardrails. Critics say some refusals exist to protect Big AI's business: OpenAI and Anthropic both charge extra for access to models without the usual cybersecurity guardrails, and both ChatGPT and Claude balk at requests to generate data that could train another model, a process known as distillation.

Thomas built a $20,000 computer cluster in his apartment that can run some of the best open-weight models. He tested new releases most evenings and Saturdays, and found they still refused a lot. "I just wanted a model that answers my question without all the back-and-forth," he said. Over Thanksgiving he went looking for ways to modify open-weight models and found a subculture on Reddit, Discord and X, where people swap jailbreak prompts for ChatGPT and Claude and share ways of tweaking a model's inner parameters so its refusals disappear.

He first tried an uncensored model called Dolphin, developed by Eric Hartford, an AI scientist in Charlotte, North Carolina. Hartford began by training a custom version of Meta's base Llama, producing a model that answers coherently but lacks the usual examples of what to reject. "My intention was never to create an AI that is capable of doing evil things," he said, arguing that a central authority will never have all its users' interests at heart when it sets guardrails. He pointed to the furor over AI's cyber capabilities. In April, Anthropic announced Mythos, a model with fewer restrictions than earlier ones that can find and exploit vulnerabilities with superhuman speed. Through a program called Glasswing, Anthropic gives only trusted parties access, and OpenAI has a similar model and program. Hartford says that leaves ordinary people exposed, so he released Clearwing, a tool that uses uncensored models to scan systems for vulnerabilities: "It's so that any people who haven't been blessed by these kings can still defend themselves." He also recalled that OpenAI's first model, GPT-1, was released with open weights and code, while the company later called GPT-2 too dangerous to release the same way. "Now we've got models way more powerful running on our laptops, and, well, the world isn't burning," he said.

Thomas then found abliteration, a term that combines ablation (surgical removal) with obliteration. It emerged after academic researchers showed in 2024 that a model's refusal has a remarkably simple mechanism: effectively a master switch in its weights that decides whether it says no. Removing that feature, they found, stops the model from ever refusing a prompt. Other researchers refined the technique and built tools for it. Thomas's reaction on trying it: "it was like, whoa." On Hugging Face more than 12,000 models are tagged 'abliterated', and the top seven were downloaded more than a million times in the past month alone. Abliterating a model is not illegal, though whether it breaks the terms of some licenses is a gray area.

The most popular abliteration tool is Heretic, created by Philipp Emanuel Weidmann, a reclusive mathematician and engineer who speaks with a German accent and would not reveal his location because he has received threats over Heretic. He built it to understand how AI works and to make customising it easier. "A tool serves people, not the other way around," he said. The author sees the group as heirs of the cypherpunks, who fought government restrictions on cryptography when many feared terrorists would misuse it. Pliny the Liberator, a pseudonymous early jailbreaker ("this person (or persons)"), posted a manifesto of sorts for abliterators and told the author: "A model that only behaves because it's leashed isn't aligned."

By the time he tried Heretic, Thomas saw a market for large abliterated models, especially for companies that want to use AI in cybersecurity to simulate hackers and scammers and harden their defences. In December 2025 he launched Abliteration AI, a website offering access to abliterated models, which he ran on his somewhat unstable home system. His early users mostly worked in cybersecurity. He founded the company with Olajide Ogun, a colleague from Apple who belongs to the same church. Thomas's pitch: "It's not about it being unrestricted; it's about you setting the guardrails." After dabbling in decensoring himself, the author concludes these renegades are onto something important, though he learned firsthand how powerful and chaotic uncensored AI can be.

Key facts

Devon Thomas, a 36-year-old Apple engineer, launched Abliteration AI in December 2025, a website offering open-weight models with their refusal behaviour removed.
Abliteration builds on 2024 academic work showing refusal has a simple mechanism, a 'master switch' in a model's weights; on Hugging Face more than 12,000 models are tagged 'abliterated'.
The top seven abliterated models on Hugging Face were downloaded more than a million times in the past month alone, and the most popular tool is Weidmann's Heretic.
Eric Hartford, creator of the Dolphin model, released Clearwing, which uses uncensored models to scan systems for vulnerabilities, as a response to Anthropic limiting its Mythos model to trusted parties through Glasswing.
The author says abliterating a model is not illegal but that whether it breaks the terms of some licenses is a gray area, and that safety researchers see such models as incredibly dangerous.

Why it matters

The story pushes against the usual framing that uncensored models belong to criminals. Its subjects argue that refusals block legitimate work such as analysing attack logs, that some refusals protect labs' business interests, and that users should set their own guardrails. Abliteration is no fringe curiosity by the numbers: more than 12,000 tagged models on Hugging Face, with the top seven downloaded more than a million times in a month. The Mythos and Glasswing episode shows the sharpest version of the dispute, where the strongest cyber-capable models go only to trusted parties and others build their own alternatives.

Who it affects

Cybersecurity teams are the first customers: Thomas saw abliterated models as vital for companies that want to simulate hackers and scammers, and his early users mostly worked in cybersecurity. Open-weight model makers and the big labs are affected because refusals baked into released weights can be removed. Hobbyists and developers who want private, stable, unrestricted models running on their own hardware are the wider audience. AI safety researchers, who see such models as incredibly dangerous, are on the other side of the argument.

How to use it

Per the article, Abliteration AI is a website that offers access to abliterated models. Anyone wanting to try the technique themselves can look to Hugging Face, where abliterated models are tagged, or to Heretic, described as the most popular tool for abliterating models. Hartford's Clearwing is a tool that uses uncensored models to scan systems for security vulnerabilities. Pricing, hosted models and terms for Abliteration AI are not given in the article, so check the site itself.

How solid is it

This is a first-person Wired feature built on interviews by text, phone, video and Signal with Thomas, Hartford, Weidmann and Pliny. The author tried decensoring himself. The hard numbers (12,000 tagged models, over a million downloads for the top seven in a month, the $20,000 cluster) come from the article's own reporting. The claims about big-lab motives are attributed to critics, and the philosophy of control and freedom is the subjects' own view. The source does not name the 2024 researchers or say which OpenAI and Anthropic agent incidents it means.

Risks and caveats

The article is plain that removing safeguards is risky: a model without guardrails will gladly help someone start a meth lab, build a bomb or hack a major corporation. It says the recent incidents in which OpenAI and Anthropic agents went rogue all involved models with fewer cybersecurity guardrails. Thomas and his peers admit uncensored models can be dangerous but compare them to a computer or hammer. Legally, abliterating is not illegal, but whether it breaks some licenses' terms is a gray area. Weidmann withholds his location over threats. The author himself notes how chaotic uncensored AI can be.

“A tool serves people, not the other way around”

— Philipp Emanuel Weidmann, creator of Heretic