I don't know. Would require serious measures and cooperation between countries, and I don't think the world is ready for that yet.

But I'd say, classifying the research and trained models, limiting the access and functionality of equipment that can be used for AI training.

Especially the more general-purpose models, like programming ones.