If you imbue a superintelligent AI with the sole goal to self-destruct, it will of course happily do so. However, the point is that it will resist being shut down if you give it any goal that it needs to remain operational to accomplish—and this covers almost all goals!
