AE Studio, Anthropic Test GRAM Across 7 Model Sizes for Switchable AI Knowledge Control
Updated
Updated · Anthropic · Jul 8
AE Studio, Anthropic Test GRAM Across 7 Model Sizes for Switchable AI Knowledge Control
2 articles · Updated · Anthropic · Jul 8
Summary
Seven tests from 50 million to 5 billion parameters found GRAM could turn four dual-use knowledge domains on or off while matching the performance of training separate filtered models.
GRAM works by adding removable modules to each transformer layer and freezing general weights on dual-use data, so knowledge in areas like virology or cybersecurity stays isolated and can later be deleted.
In larger-model experiments, removing a module cut the targeted capability about as effectively as never training on that data, without hurting general performance; post-training unlearning was easier to reverse with fine-tuning.
One GRAM training run produced 16 possible model configurations from four categories, potentially avoiding the cost of training multiple frontier models for different access levels.
Anthropic said the work is still preliminary, has not been used in any Claude production model, and may not cleanly separate dual-use knowledge that is deeply entangled with general capabilities.
Scientists can now create AI 'skill modules'. Could these modules become the next digital weapon to be stolen?
AI now has a 'delete button' for dangerous skills. But what if the deleted files can be recovered?
Switchable AI Knowledge: How GRAM Enables Robust Dual-Use Control and Safety
Overview
As AI models become more advanced, managing dual-use knowledge—information that can be used for both good and harm—has become a major challenge. Traditional methods like classifiers and refusal training often fail to control access robustly without hurting the model’s helpfulness. Gradient-Routed Auxiliary Modules (GRAM) offer a new solution by isolating sensitive knowledge in special auxiliary modules within the model. This modular approach allows precise control over what the AI can access, making it possible to switch off risky capabilities without affecting general performance. GRAM represents a promising step toward safer, more controllable AI systems.