🔓Abliteration AI Strips Safeguards From Open-Weight Models for Anyone to Use

Your local models might be refusing you for no good reason

A startup called Abliteration AI is removing safeguards from open-weight models and giving the uncensored versions away to anyone who wants them. The mainstream safety story says these models mostly serve criminals. The people building them say the refusals are the real problem. If you do security work, you've probably hit a wall. Base models refuse legitimate cybersecurity requests, and open-weight models ship with the same guardrails baked in. The founder's own account got banned after he asked a model to analyze cyber attacks. Tools like Clearwing, built on uncensored models, scan systems for vulnerabilities so ordinary people can defend themselves against hackers. The founder, 36, grew up in Indiana, studied CS at Indiana University Bloomington, and works as an engineer at Apple. He runs a $20,000 computer cluster in his apartment to host open-weight models. The Dolphin model, from a Charlotte, North Carolina AI scientist who modifies models in his spare time, responds coherently without the usual rejection examples. Worth watching: whether the rest of the industry treats refusal tuning as a product decision or a safety requirement.