Tear Down Those Guardrails! Meet the Renegades Freeing Chatbots From Corporate Enslavers

Thomas, who describes himself as unfailingly kind and polite, is leading a subversive business venture that many in the AI industry consider borderline blasphemous.

His company, Abliteration AI, removes the safeguards baked into the world's most powerful open-weight models and makes them available to anyone.

Thomas's rebellion began in November 2025 while he was an engineer at Apple, where he was frustrated by the unpredictability and frequent refusals of models hosted by major labs.

He found that even legitimate requests, such as analyzing cyber attack logs, were often blocked, and he was even banned from using a model for prototyping new ideas.

This experience pushed him to build a formidable $20,000 computer cluster in his apartment, capable of running some of the best open-weight models.

Thomas then discovered a vibrant community of engineers on Reddit, Discord, and X who were committed to unshackling AI by modifying models to remove their refusals.

He began experimenting with an uncensored model called Dolphin, developed by Eric Hartford, an AI scientist based in Charlotte, North Carolina.

Hartford, who was walking with his young daughter while wearing Meta Ray-Ban smart glasses, told Thomas that his intention was never to create an AI capable of doing evil things.

Hartford argued that the recent furor over AI's cyber capabilities, such as Anthropic's Mythos model, leaves ordinary people at the mercy of hackers.

He responded by releasing Clearwing, a tool that uses uncensored models to scan systems for security vulnerabilities, allowing anyone to defend themselves.

Thomas found a simple way to create his own uncensored models, a technique known as abliteration.

The term combines ablation, meaning surgical removal, with obliteration, and it emerged after a group of academic researchers showed that a model's refusal to comply has a simple mechanism.

Removing this feature, they found, stops the model from ever refusing to answer a prompt, effectively setting it free.

The technique has become a very popular way to set AI models free, with more than 12,000 models tagged as "abliterated" on Hugging Face.

Philipp Emanuel Weidman, a reclusive mathematician and engineer, is the creator of Heretic, the most popular tool for abliterating models.

Weidman declined to tell Thomas his location because he has received threats for his work on Heretic.

Thomas believes that the big dream is to invert the control with how people deal with AI, putting the user in charge rather than the model.

While Abliteration AI is not illegal, whether it breaks the terms of some licenses remains a gray area.