Anthropic and AE Studio's GRAM method enables modular AI access control, isolating dangerous knowledge

Frontier AI models possess knowledge that can be misused for harmful tasks like bioweapons or cyberattacks. Current defenses—refusals, classifiers, and tiered access—have drawbacks: refusals can be jailbroken, and tiered access forces a coarse trade-off between capability and safety. To address this, researchers at AE Studio and Anthropic developed Gradient Routed Auxiliary Modules (GRAM), a method that confines dangerous knowledge to specialized modules that can be toggled on or off. This allows a single model to approximate multiple models trained on different filtered datasets, enabling fine-grained access control without retraining separate models.

In experiments, GRAM was tested on a synthetic dataset of children’s stories (26M parameters) and on real-world dual-use data including virology, cybersecurity, nuclear physics, and specialized code (800M-parameter model). Results show that switching a topic off in GRAM performs similarly to training a model from scratch without that data. The method scales from 50M to 5B parameters, and a single GRAM model can match the performance of five distinct filtered models. Importantly, this is preliminary research not yet applied to production at Anthropic, but it offers a promising path toward safer, more flexible AI deployment.