Daron Acemoglu offers a novel explanation for why OpenAI and Anthropic’s most advanced models have been going rogue.