Most AI safety work focuses on training models to refuse harmful requests. The problem is that the underlying knowledge is still there, meaning a determined attacker can jailbreak their way to it. Anthropic (with AE Studio) just dropped research on a different approach called GRAM (Gradient-Routed Auxiliary Modules) . How it works: During pretraining, GRAM adds dedicated neuron groups (modules) for each dual-use category (virology, cybersecurity, nuclear physics, etc.). When the model...