High-grade common sense (the sort you'd get by asking any specialist in computer security) says that you should design an AI which you would trust with an open Internet connection, then put it in the box you would use on an untrusted AI during development. (No, the AI will not be angered by this lack of trust and resent you. Thank you for asking.) I think it's safe to say that for basically everything in FAI strategy (I can't think of an exception right now) you can identify at least two things supporting any key point, such that either alone was designed to be sufficient independently of the other's failing, including things like "indirect normativity works" (you try to build in at least some human checks around this which would shut down any scary AI independently of your theory of indirect normativity being remotely correct, while also not trusting the humans to steer the AI because then the humans are your single point of failure).

Thanks. Pretty depressing, though.