An off switch for dual use knowledge in AI models

Anthropic · 3 months ago · 44

Anthropic and AE Studio developed GRAM, a method that trains AI models with dedicated removable modules for each category of dual-use knowledge, allowing a single training run to create models configured in multiple ways without separate expensive training cycles. Testing on models ranging from 50 million to 5 billion parameters showed that deleting a module removed the corresponding capability as effectively as never training on that data, with deletion effectiveness improving at larger model sizes. This approach enables AI developers to selectively disable access to dangerous capabilities like virology or cybersecurity knowledge for different deployments while maintaining general performance.