Modular Pretraining Enables Access Control

According to the AI Alignment Forum, researchers at Anthropic and AE Studio have developed a new technique for fine-grained access control in AI models. Called Gradient-Routed Auxiliary Modules—or GRAM—it compartmentalizes dangerous knowledge, like advanced virology or cybersecurity capabilities, into separate model components that can be switched on or off. Traditional safeguards like refusals and classifiers sit on top of knowledge the model already possesses and can be jailbroken. GRAM is different: it allows a single model to be deployed with different capabilities depending on user trust levels. A vetted biosecurity lab could access advanced virology knowledge while general users couldn't—without needing separate models for each variant. Tests on models up to five billion parameters showed GRAM approximated the performance of models trained separately on filtered data. The researchers note this is preliminary work not yet applied to production systems.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton