Anthropic Tests GRAM Access Control for Dual-Use Knowledge

Anthropic's GRAM research matters because it targets a hard safety engineering question: how to limit access to dual-use model knowledge without training and operating many separate frontier models. In a July 8, 2026 post with AE Studio collaborators, Anthropic described Gradient-Routed Auxiliary Modules, a pretraining method that routes categories such as virology, cybersecurity, nuclear physics, and specialized code into removable modules. The lab emphasized that the work is preliminary and has not been applied to Claude production models. For practitioners, the useful signal is that access control may move below refusal layers, into model architecture and training data flows.