Google DeepMind researcher @iamtrask says the best way to build an AI-proof sandbox may be to train models to break it, fix it, then break it again until no holes remain: "How do we get rigorous sandboxes where the models only know the information we want them to know, so they can't work around and try to cheat, in the worst case stuff like the Hugging Face incident?" "There's a clue to how to solve really hard sandboxes in this morning's release from OpenAI. For any task where you can…