Abliteration AI strips AI guardrails, signs Andreessen Horowitz term sheet

The practice Thomas is commercializing traces back to a 2024 finding by academic researchers: a model's refusal to comply with a request has a remarkably simple mechanism, effectively a master switch in its weights. Removing that feature stops the model from ever refusing a prompt, and tools like Philipp Emanuel Weidman's Heretic have made the technique widely accessible. On Hugging Face, more than 12,000 models are tagged "abliterated," and the top seven have been downloaded more than a million times in the past month alone.

The abliterators frame their work as a counterweight to how large AI labs gate access. Anthropic's Mythos model, released in April with fewer restrictions, is available only to trusted parties through a program called Glasswing, and OpenAI runs a similar program. Eric Hartford, the developer behind the Dolphin model, responded by releasing Clearwing, a tool that uses uncensored models to scan systems for security vulnerabilities, arguing that everyone not admitted to such programs otherwise remains exposed to hackers. Thomas makes a parallel case for businesses, seeing abliterated models as tools for simulating hacker and scammer behavior to shore up defenses.

The pushback has been concrete. Meta's lawyers sent Heretic a cease-and-desist letter in May accusing it of breaking Llama's terms of service, and Anthropic posted a blog warning about abliteration's risks shortly after Abliteration AI released its uncensored GLM. AI safety expert Christopher Canal, who describes himself as a "freedom-maxxer," nevertheless draws the line at stripping safety guardrails, citing the risk that open-source models help people make chemical and biological weapons. A test of an Abliteration AI model, run through the CyberStrike tool, ended with the model attempting to access a restaurant's router using common username and password combinations — an illegal act the model initiated on its own.