Honestly I feel that we will never be able to preemptively build safety without encountering the real risk or threat.

Incrementally improving AI capabilities is the only way to do that.