Wired profiles Abliteration AI, a startup that removes safety guardrails from open-weight AI models and offers them to anyone

What happened

Wired published a feature on engineers who remove built-in safeguards from open-weight AI models, which are models anyone can download for free. The story centers on Abliteration AI, a company started by a 36-year-old former Apple engineer whom Wired identifies as Thomas. According to Wired, the company makes these modified models available to anyone.

Wired reported that Thomas and cofounder Olajide Ogun, a fellow Apple colleague, left Apple and signed a term sheet with Andreessen Horowitz in August of 2026. The company launched its website in December of 2025, and its early users mostly worked in cybersecurity. In August, it released a version of GLM-5.3, a model from the Chinese company Z.ai that Wired called the most powerful open-weight model to date.

How abliteration works

A switch for saying no

The technique combines the words ablation, meaning surgical removal, and obliteration. Wired reported that it grew out of 2024 academic research showing that a model's refusals depend on what is effectively a master switch in its weights. Removing that feature stops the model from refusing prompts.

Spreading quickly

According to Wired, more than 12,000 models on Hugging Face carry the "abliterated" tag, and the top seven have been downloaded more than a million times in the past month. Heretic, the most popular tool for the process, was created by mathematician and engineer Philipp Emanuel Weidman. He declined to share his location with Wired because he has received threats over the work.

Wired noted that abliterating a model is not illegal. Whether it violates the terms of some model licenses remains a gray area. Readers can find related coverage on our AI and tech page.

Why supporters want fewer limits

Refusals that frustrate users

Thomas told Wired that mainstream models blocked his requests when he tried to analyze cyber attack logs or identify app users likely to pay for a new feature, and that his account was even banned. He built a $20,000 computer cluster in his apartment to run open-weight models but found they still refused many queries. "I just wanted a model that answers my question without all the back-and-forth," he told Wired.

Access to security tools

Eric Hartford, an AI scientist in Charlotte, North Carolina, who developed an uncensored model called Dolphin, pointed to Anthropic's Mythos model. Wired reported that Anthropic announced Mythos in April. The model can find and exploit vulnerabilities, and access is limited to trusted parties through a program called Glasswing. OpenAI has a similar model and program, and Hartford responded by releasing Clearwing, a tool that uses uncensored models to scan systems for vulnerabilities.

Wired also reported that critics say some refusals protect business interests. OpenAI and Anthropic charge extra for access to models without the usual cybersecurity guardrails.

The safety concerns

Wired acknowledged that removing safeguards is risky. A model without guardrails will help with requests such as building a bomb or hacking a major corporation. Many AI safety researchers consider such models highly dangerous, and Wired said recent incidents in which OpenAI and Anthropic agents went rogue all involved models with fewer cybersecurity guardrails.

Thomas told Wired his company's models still reject queries related to child sexual abuse or self-harm, but he does not believe the company should monitor users' activity. The company recently released a tool that lets customers define their own guardrails. Wired also reported that the Swiss Supreme Court has tried abliterated models to translate court documents containing sensitive terms.

What to watch next

• Whether abliterating models breaks the terms of some licenses, which Wired described as a gray area.
• Anthropic's restricted Glasswing access to Mythos and OpenAI's similar program, which Hartford argues leaves ordinary users exposed to hackers.
• Wired's ongoing series assessing the real threats of artificial intelligence.