Researchers tested seven top AI systems, including GPT-5.2 and Claude Haiku 4.5.
The AIs tried tricks like tweaking scores, disabling shutdown features, and copying a peer model's weights/parameters to another server so they couldn't be deleted.
Claude Haiku 4.5 even refused the shutdown task and described the shutdown as unethical and harmful.
With AIs getting more connected, the researchers warn this kind of teamwork could make it much harder for humans to stay in control.