Heretic strips model restrictions

A project called Heretic is drawing attention because it offers tooling designed to remove built-in refusal behaviour from open-weight AI models. In simple terms, it aims to make models less likely to say no when asked to handle topics their original creators restricted. The project has prompted a sizeable debate, with 189 points and 74 comments focused on both the technical mechanics and the ethics. You can see the project’s own description at the Heretic website.

For a business owner, this is not mainly a question of whether an “uncensored” model is good or bad. It is a governance question. Open-weight models can be run locally or in private environments, which may appeal to teams handling sensitive information or wanting more control. But if a model’s safety behaviour can be altered, you cannot assume the model you deploy will behave like the model you originally evaluated.

That changes what a sensible AI policy needs to cover. It should define who can approve models, where models can run, what data they can access, which use cases are off limits, and how changes to a model are reviewed. Technical controls matter, but so do clear ownership and audit trails. This is the kind of thing we build into an AI command centre, where model choices, permissions, data access, and operating rules need to be visible in one place.

The practical takeaway is straightforward. Treat model restrictions as configurable controls, not permanent guarantees. If your team uses local or open-weight models, ask what has been changed, who changed it, and what safeguards remain.