Anthropic introduces GRAM to isolate dual-use knowledge in AI models

Anthropic, in collaboration with AE Studio, published research on GRAM (Gradient-Routed Auxiliary Modules).

GRAM adds dedicated, removable neural modules for categories of dual-use knowledge such as virology, cybersecurity, and nuclear physics.

During training, only the relevant module learns from sensitive data; the module can later be deleted to remove the capability without retraining the whole model. The work is preliminary and not yet used in production models.