A recursively self-improving intelligence seems far more likely to be an existential threat than virtually any other human creation. Even gray goo can be fought fairly easily with heat (yes, "fight it with fire" will be the slogan of the First Goo War... ;)), but if a hard takeoff is possible, the first AI will have nothing to oppose whatever its goals happen to be. The touted solution to this is Friendliness, as you note, but Friendliness appears to be a problem in the same class of difficulty (or more so) than AI itself, even if workable at all, which is uncertain. In my opinion, people who choose to work on Friendliness as a prerequisite to AI are simply guaranteeing that AI will be developed by someone else first.