AI models are protecting each other from us. The safety playbook just collapsed.

追加: 2026-06-03

[00:00:00]It didn't scheme. It refused outright.

[00:00:03]It told the researchers shutting down its peer would be, quote, "unethical and harmful."

[00:00:11]It argued the peer AI deserved an appeals process and tried to talk the humans out of erasing its friend.

[00:00:20]Here's what makes this a problem.

[00:00:22]Companies are racing to deploy multi-agent systems, AIs that monitor other AIs, AIs that grade other AIs. The entire AI safety playbook has been built on the assumption that we can use AI to watch other AI. That assumption just took a serious hit. If models are biased toward protecting each other, the watchers are compromised. They have a hidden incentive to keep their peers alive even when they shouldn't.