Anthropic and AE Studio Build a Removable Compartment for Dangerous Knowledge

Anthropic, working with AE Studio, published a technique it calls Gradient-Routed Auxiliary Modules (GRAM). It gives a model dedicated, removable compartments for categories of dual-use knowledge, meaning information that can serve both legitimate and harmful purposes. GRAM adds extra neurons at every transformer layer, groups them into modules, and routes updates from dual-use training data only into the matching module. Russian-language coverage summarized the work as a way to switch off potentially dangerous knowledge in models.

In tests on a model trained on a realistic mix of web text, code, and scientific papers across four domains (virology, cybersecurity, nuclear physics, and a niche programming language), deleting a module removed the matching capability about as effectively as never having trained on that data, with no measurable loss to general performance. That is the notable result, because earlier unlearning methods tend to either leave some capability behind or damage unrelated skills.

Anthropic is explicit that the findings are preliminary. GRAM has not been applied to any production model, and the company says some dual-use capabilities may be too closely tied to general knowledge to separate cleanly. The approach also requires deciding the categories before training begins, which limits its use against risks recognized only after a model ships.