The gist: GRAM partitions dual-use knowledge (such as virology or cybersecurity) into dedicated, removable neuron modules, allowing a trained model to be flexibly configured for different security requirements without needing to train separate models.
Anthropic and AE Studio have developed a method called GRAM that makes it possible to isolate and control dangerous dual-use knowledge in AI models in separate, exchangeable modules. This could make it possible to configure a single trained model for different security profiles.
Frontier AI models store large amounts of knowledge, including so-called dual-use knowledge – capabilities that can be used for both constructive and harmful purposes. Examples include cybersecurity knowledge for patching security vulnerabilities or exploiting them, as well as virology knowledge for vaccine development or the construction of pathogenic substances.
Previous security measures such as training refusals and input/output filters fall short: they do not change the knowledge stored in the model itself, so sufficiently determined attackers could access the dual-use knowledge through jailbreak techniques. A more robust solution lies in controlling what the model knows in the first place. Until now, filtering training data or isolation in “removable slices” was impractical because a separate, highly costly model had to be trained for each different configuration.
GRAM (Gradient-Routed Auxiliary Modules) solves this dilemma: the method adds additional neurons to each layer of a transformer, which are organized into modules per dual-use category. During training, these modules are only activated and updated when the model encounters dual-use data (e.g., virology texts). The general weights remain frozen. The dual-use knowledge thus concentrates in its specialized modules rather than being distributed across the entire network. After training, the modules can simply be deleted or activated for trustworthy deployments. A single trained model can thus be flexibly configured: with four dual-use categories, 16 different configurations emerge (per category “on” or “off”).
Anthropic emphasizes that the preliminary results to date are preliminary. GRAM has not yet been deployed in Anthropic’s production models, and it is uncertain whether this will happen in the future.
Source: www.anthropic.com · Published 8 July 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.7.3.