Right, but the reward-hacky nature of these models calls into question their usefulness as cyberweapons.

How can you trust it when it goes "I superhacked the Chinese servers as you requested, and here are the classified documents which I definitely didn't fabricate."