I know the architecture of Deep Mind, (It's reinforcement learning + deep learning, basically) and can guarantee you that 1000x computing power would have a hard time getting you to NES games, let alone anything dangerous.

As long as the computer is in its own simulated world, with no input from the outside world, we're almost certainly safe. It cannot model the real world. But hook it up to some cameras and microphones, and then you have the potential for something that could wind up being dangerous. So I'd say there's no reason to speculate about 1000x computing power. Just stick it in a virtual world with no human communication and let in run for a while and see if it shows signs of the kind of intelligence that would be worrying. (The AI Box argument does not apply here) The challenge, of course, is coming up with a virtual world that is complex enough to be able to discern high intelligence while being different enough from the real world that it could not apply knowledge gained in the simulation to the real world.