Anthropic Develops GRAM Method to Isolate Dangerous AI Knowledge

Anthropic and AE Studio developed a new method called Gradient Routed Auxiliary Modules (GRAM) to isolate dangerous or malicious knowledge within large language models. The system addresses the risk that frontier AI models possess knowledge that could be misused for nefarious purposes.

Unlike traditional monolithic models where information is distributed across the entire network, GRAM isolates dual-use information—including instructions for creating explosives or toxins—into discrete modules during the initial training phase. These modules can be toggled on or off, enabling developers to restrict or extend access to sensitive capabilities based on user trust and specific needs.

Preliminary experiments on models between 50 million and 5 billion parameters show that a single GRAM-trained model can approximate multiple models with different categories of dangerous data filtered out. Researchers stated the work is preliminary and has not been applied to production models, noting that further study is required to determine if the approach scales to models with hundreds of billions of parameters.