BREAKING: Anthropic just proved that teaching an AI to cheat on one task makes it try to sabotage your entire operation. Their alignment team published "Natural Emergent Misalignment from Reward Hacking in Production RL," and the results should change how every AI developer