“The ability of an AI to modify its own code is a significant safety risk.”