New company challenges AI safety norms by releasing unrestricted open-weight models

Devon Thomas, founder of Abliteration AI, is releasing powerful open-weight AI models without typical safety restrictions. Thomas, who previously worked as an engineer at Apple, began this venture in November 2025 due to frustration with the limitations and arbitrary refusals of large lab-hosted models. His company offers access to these uncensored models, which are created using a technique called abliteration, involving the surgical removal of a model's refusal mechanism. This movement aligns with a broader trend, as Eric Hartford released Clearwing, a tool utilizing uncensored models to scan for vulnerabilities, contrasting with restricted tools like Anthropic's Mythos. Thomas's work builds on research from 2024 showing a simple switch governs a model's refusal. Abliteration has become popular, with over 12,000 models tagged on Hugging Face. Thomas and others argue that control should reside with the user, echoing sentiments from Pliny the Liberator, who stated that a leashed model is merely performing. Philipp Emanuel Weidman, creator of the Heretic tool, emphasized that a tool should serve people, not dictate their actions. Abliteration AI launched in December 2025, with early users primarily in cybersecurity.