An off switch for dual-use knowledge in AI models

Anthropic Research, in collaboration with AE Studio, has detailed preliminary research on a method called GRAM, which stands for Gradient-Routed Auxiliary Modules. The technique provides AI models with dedicated, removable compartments for categories of dual-use knowledge, such as cybersecurity, virology, nuclear physics, and specialized code. Standard safeguards like refusal training and input classifiers only prevent dangerous outputs without altering the underlying knowledge, and filtering out data during pretraining requires training multiple separate models at a high cost. GRAM addresses this by adding extra neurons to standard Transformer layers that are grouped into modules. When a model encounters dual-use text, only the corresponding module is allowed to learn from it while general-purpose weights are frozen. After training, the specific module can be deleted to remove the capability or retained for trusted users. In experiments across model sizes ranging from 50 million to 5 billion parameters, GRAM matched the performance of data filtering at the cost of training a single model, and removing modules did not degrade general performance. Anthropic notes that the research is still preliminary, GRAM has not been applied to production models like Claude, and evaluations focused on next-token prediction rather than real downstream tasks. Furthermore, some dual-use capabilities may remain deeply entangled with general knowledge.