CBRN risk mitigation with Anthropic

our joint research on GRAM is now out

July 14, 2026

Hey AE Studio Bytes-ers,

Our joint research with Anthropic is now out on their Alignment Science blog and the ICML 2026 spotlight paper can be seen here.

The research tackles the problem underneath basically every AI safety headline. Models learn dangerous knowledge during training because it's tangled up with useful knowledge; you can't teach a model biology without brushing up against virology. The standard fix is to train the model to refuse and add classifiers to catch bad requests. But the knowledge is still in the model, which is why jailbreaks keep working. You're not removing the dangerous capability, you're asking the model nicely not to use it.

GRAM (gradient-routed auxiliary modules) takes a different approach. During pretraining, we route the dangerous knowledge into small removable modules, one per domain, so it never gets absorbed into the model's general weights. Delete a module at deployment time and you get a model that performs almost identically to one that was never trained on that topic in the first place.

Another part we’re excited about: when we simulated attackers trying to fine-tune the knowledge back in, the removal held. We ran the same attack against a leading unlearning method and the knowledge came back to near baseline, which suggests unlearning suppresses capabilities rather than removes them. GRAM's advantage also grew with model scale (tyvm for compute, Anthropic!).

There's a practical upside too. Normally, if you wanted five versions of a model with different capability sets, you'd train five models. GRAM gives you all of them from one training run: full capability for a vetted research lab, verifiably absent everywhere else, just by deleting weights.

The usual caveats apply. This is early research, tested up to 5B parameters, and not yet applied to production models. But if safety can live in what a model is rather than in how it's been asked to behave, that changes what enterprises, regulators, and anyone putting AI in front of people can actually verify.

Read the technical post, Anthropic's plain-English version ("An off switch for dual use knowledge in AI models"), and grab the code, all public.

Or if you prefer to listen, you can also hear more about it from our own researchers, breaking it down on our latest podcast episode.

——————————————————————————————————————————

AI Alignment is the research field working on a single question: how do you build AI systems that stay beneficial as they become more capable? Our approach treats alignment as something you build into a system from the start, the way evolution built prosociality into humans through empathy, honesty, and accurate self-models. These properties turn out to correspond with capability gains, so alignment done right makes models smarter and safer at once. We call this the negative alignment tax, and we think solving alignment is the most important engineering problem in the world. If you’re new to AI alignment, start with our introduction to AI safety and alignment here.

AE Studio pursues neglected approaches like gradient routing and self-other overlap, and our consulting business focuses on high-stakes production AI and funds the research with no strings attached.