The reason most AI safety ideas break, at least in my mind, is that when looking at longer horizons AI can build AI. Well, that and unaligned humans too. If all that's keeping us safe is that AI will off itself, then the first moment non-offing AI shows up we're in trouble.