Devon Thomas hardly seems like a renegade to members of his bible study group, who describe him as unfailingly kind and polite. Yet Thomas is spearheading an initiative that many in the artificial intelligence industry consider subversive: his startup, Abliteration AI, removes safety safeguards from the world’s most powerful open-weight models and makes them freely available to anyone. A 36-year-old former Apple engineer who studied computer science at Indiana University, Bloomington, Thomas launched the venture in December 2025 alongside his colleague Olajide Ogun. His tech rebellion began in November 2025 out of personal frustration, after mainstream models repeatedly blocked his requests to analyze cyber attack logs or determine which users might pay for new features.
Mainstream systems developed by major labs are trained to refuse a wide range of queries, including legitimate cybersecurity tasks, adult topics, and politically sensitive subjects. While open-weight models ship with similar guardrails baked in, critics point out that these strict limitations often serve to protect corporate business models. OpenAI and Anthropic charge extra for access to less-restricted models and reject requests for data distillation. Seeking complete control, Thomas constructed a $20,000 computer cluster in his apartment to test free-to-download models in his homelab. That search led him into a vibrant online subculture on Reddit, Discord, and X, where engineers share jailbreak prompts and techniques to modify model parameters.
Key takeaways
- Devon Thomas and Olajide Ogun launched Abliteration AI in December 2025 to provide access to safeguard-free open-weight models.
- Abliteration targets a specific master switch within an AI model's inner weights to remove its ability to refuse prompts.
- Over 12,000 models are tagged as abliterated on Hugging Face, with top versions surpassing a million downloads in a single month.
- Independent developers like Eric Hartford and Philipp Emanuel Weidman have created tools such as Clearwing and Heretic to facilitate AI customization.
The core practice of stripping these restrictions relies on a technique known as abliteration, combining ablation and obliteration. Stemming from 2024 academic research, the method demonstrates that a model's refusal mechanism is controlled by a simple master switch within its weights. Removing this feature stops the model from ever saying no to a prompt. Philipp Emanuel Weidman, a reclusive mathematician and engineer who communicates via Signal, created Heretic as the most popular tool for abliterating models to help users customize AI. Similarly, AI scientist Eric Hartford developed Dolphin by tuning Meta's base Llama model, and later released a tool called Clearwing to help everyday individuals scan systems for security vulnerabilities without relying on restricted corporate platforms like Anthropic's Mythos.
While safety researchers warn that removing these guardrails can help bad actors build bombs or hack major corporations, proponents argue that uncensored models simply put the user in charge of their own security. Supporters draw direct parallels to the cypherpunk movement of the internet's early days, viewing abliteration as a vital fight for cognitive freedom and privacy. Thomas ultimately sees his abliterated models as essential resources for corporate cybersecurity teams, allowing them to accurately simulate the behavior of hackers and scammers to effectively shore up their defenses.
Comments