Anthropic and AE Studio publish GRAM (Gradient-Routed Auxiliary Modules) research on the alignment blog, proposing a brand-new "architecture-level" safety access control approach. It attaches small auxiliary modules in parallel to each MLP layer of the Transformer, routes gradients by data type during training, and removes the corresponding modules at inference to shut off specific capabilities — without needing to retrain the whole model. Experiments span 50M to 5B parameters, successfully isolating four categories of dual-use knowledge — virology, cybersecurity, nuclear physics, specialized code — into independent modules; a single GRAM model can reconstruct five different filter configurations, and combining 4 modules yields 16 switch states. On dual-use capability preservation and forgetting, adversarial fine-tuning robustness, compositionality, and partial-annotation scenarios, GRAM outperforms post-hoc forgetting methods like MaxEnt and LoRA fine-tuning baselines. This is the first time access control on frontier models has been pushed from the "refuse training + classifier" behavior layer to the "weight topology" structure layer. Anthropic emphasizes the work has not yet entered production Claude, but the line of thinking is worth tracking long-term: in the future, the same base model might be able to dynamically slice by user trust level, "lighting up on demand, shutting down on demand" — provided it scales to hundreds of billions of parameters and resolves open issues like instruction-tuning compatibility and entangled-capability separation.