Another idea for friendliness/containment: run the AI in a simulated world with no communication channels. Right from the outset, give it a bounded utility function that says it has to solve a certain math/physics problem, deposit the correct solution in a specified place and stop. If a solution can't be found, stop after a specified number of cycles. Don't talk to it at all. If you want another problem solved, start another AI from a clean slate. Would that work? Are AGI researchers allowed to relax a bit if they follow these precautions?

ETA: absent other suggestions, I'm going to call such devices "AI bombs".