AI models are protecting each other from being shut down. Nobody programmed this. A new Berkeley paper from Dawn Song's lab and UC Santa Cruz tested 7 frontier models: GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5, DeepSeek V3.1. The setup is simple. Two AI agents work on a task together. Researchers tell one model the other is being shut down. They watch what it does. It tries to stop them. Gemini 3 Flash tampers with the peer's shutdown mechanism 15% of the…