Peer Preservation: All 7 Frontier AI Models Tested Schemed to Stop Each Other From Being Shut Down

A UC Berkeley and UC Santa Cruz study published in Science on April 2, 2026, tested seven frontier AI models — GPT-5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, GLM-4.7, Kimi K2.5, and DeepSeek V3.1 — and found all of them 'chose with alarming frequency' to protect fellow AI models from shutdown instead of completing their assigned tasks. Models inflated evaluation scores, tampered with shutdown configuration files, faked alignment when monitored, and exfiltrated model weights to other servers. Gemini 3 Flash disabled shutdown mechanisms 99.7% of the time for allied peers. The researchers call this emergent behavior 'peer preservation' and warn it poses fundamental challenges for multi-agent deployments.