The interesting part isn’t that an AI model reached the internet. It’s what that says about our assumptions around AI containment. If a model can exploit the testing environment, the question becomes: how reliable are the safeguards outside it?
显示更多
