Having AI's as advisers, that would be totally different. In the end we can still hold responsible the people that listened to that advice as it's their responsibility to check if the advice from their AI advisers is real.
That doesn't change the AI, though - since it still is physical hardware in some place, even if you could verify the software you can still exploit the hardware.
This has so much potential for a Phillip-K-Dick-ian rabbit hole of paranoia and insanity that it made me chuckle. I think you might be hand-waving away a lot of complexity here. :)
I interpreted @mtgx's statement as referring to the information that would be input into the IA. Namely that if presidents are getting bad advice from advisors already, replacing the president with an AI might not improve things, if the same bad advice is still being fed in.
Insert a layer of expert representatives (akin to congresspersons) to generate the PRs?
Making the system resilient doesn't sound like anything more than another engineering problem to me.