Anthropic's New Technique Puts a Delete Button on Dangerous Model Knowledge

Anthropic and AE Studio built a technique that lets a language model’s dangerous knowledge be surgically removed after training, without retraining the whole model. Gradient-Routed Auxiliary Modules (GRAM) add dedicated neurons to every Transformer block’s MLP layer for each dual-use category — virology, cybersecurity, nuclear physics, specialized code — and route only the relevant modules into the forward and backward pass during training on sensitive data. Deleting a module afterward removes the associated capability almost as completely as if the model had never seen the training data, while general performance stays close to baseline. Tested across model sizes from 50 million to 5 billion parameters, GRAM matched the performance of straightforward data filtering at every scale, with the gap between “module on” and “module off” widening as models got larger.

It landed two days after Anthropic’s J-lens interpretability paper, which locates the small internal workspace where a model’s reasoning becomes visible. The two results are complementary halves of the same problem: J-lens finds where reasoning happens inside a model; GRAM controls where specific knowledge lives well enough to strip it out later. Neither is about making a model behave better in the moment — both are about giving a lab a handle on the model’s internals it didn’t have before.

Anthropic is unusually direct about where this stands: the results are preliminary, GRAM hasn’t been applied to any production model, and the company says it isn’t certain it ever will be. That hedge is the useful signal for anyone evaluating an AI vendor’s safety claims. A research paper describing a capability and a shipped control that a company actually relies on are different things, and the gap between them is exactly where due diligence should focus — ask what’s running in production, not what’s been published.