Anthropic Introduces GRAM Method to Isolate Dangerous Knowledge in AI

Anthropic researchers have developed the GRAM (Gradient-Routed Auxiliary Modules) method, which allows for the physical isolation or removal of "dual-use" knowledge from neural networks.

Compiled by Sergey KostenchukPublished 2026-07-09Updated 2026-07-09

2026-07-09 Safety Anthropic

🛡 Anthropic Introduces GRAM Method to Control Dangerous Knowledge in AI

Researchers have developed the GRAM (Gradient-Routed Auxiliary Modules) method, which allows for the isolation of "dual-use" knowledge into separate modules within a neural network. This enables the creation of a single model that can be customized: keeping dangerous knowledge available for vetted laboratories or completely removing it for public use. Unlike standard refusal training, GRAM physically removes knowledge from specific neurons, making it practically irrecoverable even if the model is hacked.

🌍 The method addresses the high cost of creating multiple specialized models. Instead of training different versions, a single architecture with module "switches" can be used, providing precise access control to sensitive data.

👤 This is a step toward safe AI systems that remain useful to scientists without becoming tools for creating biological or digital weapons in the hands of malicious actors.

Source 1: https://www.anthropic.com/research/off-switch-dual-use

Sources