Well the first thing we can do is start disconnecting safety critical infrastructure from the Internet and/or radio networks... This stuff should never have been online in the first place.
I in no way want to seem like I endorse EY's brand of crackpotery, but do note he says in the interview a malicious AI would go undetected and would not try persuading humans to do anything to "get out of jail"; it would instead hack the system. EY seems to think a hostile AI would not contact human proxies in order to start its nefarious plans, because we'd be too slow for it, and also because alerting any of us would be a risk.
For getting "out of jail", sure. But considering how we're already connecting it to the internet, and I doubt that is going away, it seems like a moot point. For other nefarious plans, I think it can contact human proxies without anyone thinking it was actually an AGI. Perfectly capable of sending emails and signing it as a John.
I don‘t think so, instead there should be a very simple fixed-width formally proven protocol per use case over a very basic bus connected to an internet gateway.
