I haven't seen the livestream, but I just heard that they intend to have their AI automatically change your passwords on websites if it considers them insecure, which sounds to me like the worst idea for AI so far, and that bar is high.
For the vast majority of services, even if this action fails and the wrong password is saved (!?) you're still just a "forgot password" click away.
By contrast, an LLM can do it easily.
But that's all of them though?