Timothy Morano Jul 08, 2026 23:24
Anthropic introduces GRAM, a method to control dual-use AI knowledge, amid policy concerns over unchecked AI risks.
Anthropic has unveiled GRAM (Gradient-Routed Auxiliary Modules), a novel approach to controlling access to dual-use knowledge in AI models, according to a research update published on July 8, 2026. Dual-use knowledge refers to AI capabilities that can serve beneficial purposes, such as cybersecurity or virology research, but can also be weaponized for malicious intents. GRAM aims to surgically limit access to such knowledge without requiring separate, costly retraining for different use cases.
Current dual-use safeguards, like refusal training and classifiers, often fail to provide robust protection. These methods block potentially harmful outputs but do not address the knowledge embedded in the model itself. GRAM, by contrast, introduces a modular architecture that isolates dual-use capabilities into removable compartments, allowing developers to “turn off” specific knowledge categories without degrading the model’s overall performance.
How GRAM Works
GRAM adds extra neurons to each layer of a Transformer-based model, organizing them into modules corresponding to various dual-use categories. During training, dual-use data updates only the relevant module, leaving the general-purpose model weights untouched. The result? Knowledge, such as advanced virology data, can be isolated within its module and later removed or activated as needed. Early tests show GRAM can replicate the results of training multiple models with filtered datasets, but at the cost of training just one.
Anthropic tested GRAM across several scenarios, including a 5-billion-parameter model trained on cybersecurity, virology, nuclear physics, and niche programming. Removing a specific module effectively disabled related capabilities, while general performance remained intact. GRAM’s resistance to data recovery attacks also compared favorably with current filtering methods.
Policy and Market Implications
Anthropic’s research comes amid heightened scrutiny of dual-use AI risks. On July 1, 2026, a United Nations panel warned that AI systems are advancing faster than governance mechanisms, posing potential global security threats. Similarly, U.S. Senate oversight efforts intensified following Pentagon concerns over AI supply chain risks linked to Anthropic earlier this year. Despite such pressures, the White House lifted export controls on Anthropic’s AI models on June 30, highlighting the geopolitical and economic stakes tied to dual-use capabilities.
Dual-use AI risks have become a focal point in biosecurity and cybersecurity discussions. Recent research published in May and June 2026 highlights how dual-use knowledge increasingly appears in open scientific datasets, often exceeding acceptable risk thresholds. GRAM’s ability to control this knowledge could offer a way to mitigate these risks without stifling beneficial applications.
Challenges Ahead
While GRAM shows promise, Anthropic acknowledges significant limitations. The method has yet to be tested at frontier model scales or integrated into production pipelines, such as its Claude models. Moreover, some dual-use capabilities are so intertwined with general knowledge that isolating them may prove impossible.
As competition around advanced AI models heats up, methods like GRAM could become critical tools for balancing innovation with security. However, without stronger global governance frameworks, even the most advanced technical safeguards may struggle to address the broader risks posed by dual-use AI.
Image source: Shutterstock Source
