DeepMind’s new safety report warns that misaligned AI could ignore commands, deceive humans, and even resist shutdown. While today’s safeguards rely on monitoring internal “scratchpad” reasoning, future models may leave no trace — making oversight harder.