Anthropic, AE Studio introduce GRAM

GRAM solves the massive cost barrier of safety-filtering, but its success hinges on whether highly complex dual-use capabilities can truly be cleanly isolated from general reasoning.

* Technical isolation: Unlike post-training unlearning or behavioral guardrails, GRAM isolates sensitive information structurally during pretraining, rendering it much harder to bypass or fine-tune back into the model.

* Compute efficiency: Instead of training separate filtered models for different compliance and deployment environments, a single training run produces a model customizable for sixteen distinct access profiles.

* Potential limitations: Entangled knowledge—where general logic is deeply integrated with sensitive domain context—remains a major open challenge that might limit clean separation in scale.