Anthropic Collaboration Unveils GRAM, a Knowledge Access Control Technique

This paper introduces GRAM (Gradient-Routed Auxiliary Modules), a new approach to solving the dual-use dilemma of AI models—the problem of simultaneously possessing beneficial knowledge (e.g., vaccine design) and dangerous knowledge (e.g., pathogen design).

Existing refusal learning or data filtering methods have limitations, being either vulnerable to jailbreak attacks or extremely costly. GRAM takes the approach of pre-designating an 'address (auxiliary module)' where a specific category of knowledge will be stored, starting from the training stage, before the knowledge spreads throughout the neural network.

Key Features and Advantages:

• Modular Access Control: By attaching category-specific Auxiliary Modules to the transformer's MLP blocks, specific knowledge modules can be turned on or off depending on the deployment target's permissions.
• Efficiency: A variety of Capability Profiles can be supported with just a single training run, without needing to repeat pretraining multiple times.
• Security: Because the knowledge itself is isolated within specific modules of the model, it provides far stronger defense against jailbreaks and adversarial attacks than simple modifications at the behavioral layer.

This research has been accepted as an ICML 2026 Spotlight, presenting a next-generation AI safety technique that can precisely control access to dangerous knowledge while maintaining model performance.