Re: AIs in a simulation, it seems like whatever goals the AI had would be defined in terms of the simulation (similar to how if humanity discovered we were in a hackable simulation, our first priorities would be to make sure the simulation didn't get shut off, invent immortality, provide everyone with unlimited cake, etc.--all concerns that exist within our simulation.) So even if the AI realizes its in a simulation, having its goal defined in terms of the simulation probably counts as a weak security measure.