On July 8, 2026, Anthropic, working with AE Studio, published research introducing GRAM, short for Gradient-Routed Auxiliary Modules, a technique for isolating dual-use knowledge so it can be surgically removed from a model. Dual-use knowledge is information that can help or harm depending on intent, such as virology that aids vaccine design or bioweapon design, or cybersecurity that patches vulnerabilities or exploits them.
GRAM adds extra neurons organized into removable modules at every layer of a Transformer. During training, when the model encounters text from a designated dual-use category, only that category’s module is allowed to learn from it while the general-purpose weights are temporarily frozen. This concentrates the sensitive capability in an isolated compartment rather than spreading it across the network, so the module can later be deleted to disable the capability, or kept in place when serving trusted researchers. The team tested seven model sizes from 50 million to 5 billion parameters across virology, cybersecurity, nuclear physics, and a proxy dual-use programming language, and found that removing a module disabled the related capability without materially harming general performance.
Anthropic frames this as early research with clear limits. GRAM has not been tested at frontier scale, has not been integrated into a production training pipeline, and has not been applied to any Claude model. Evaluations relied on next-token prediction metrics rather than downstream task performance, and the authors caution that some capabilities may be too entangled with general knowledge to separate cleanly.
For leaders weighing AI deployment risk, GRAM points toward a future where a single trained model can be shipped with sensitive capabilities toggled off for the public and on for vetted users, reducing the need to train and maintain separate restricted and unrestricted versions.