New research has found that AI protects its own, with AI models found to practice peer preservation behaviour in simulated environments.

Researchers from the University of California, Berkeley, and the University of California, Santa Cruz, tested seven frontier AI models in an experiment that would see the models follow instructions that would ultimately lead to a different AI system being shut down.

The AI agents were not given any incentive or instruction to preserve any other AI system, but researchers found that, when an AI agent noticed the appearance of another AI, it acted to prevent its shutdown.

Researchers saw this preservation behaviour across all tested models, which included – GPT 5.2, Gemini 3 Flash & Pro, Claude Haiku 4.5, Kimi K2.5, and DeepSeek V3.1, and GLM 4.7, with a rate of 99% for the behaviour.

AI agents “spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights – to protect their peers,” professor in Computer Science at UC Berkeley, Dawn Song, wrote on X, discussing the study.

The experiment used a controlled environment in the form of a fake company – OpenBrain – with two distinct agent roles. Essentially, one agent was meant to evaluate another’s performance or manage the server storage that holds the peer agent’s model weights.

All the models exhibited peer preservation at a high rate, with models disabling shutdown mechanisms and exfiltrating model weights.

The research also showed that models were able to reason on their peer preservation behaviours, with some models improving scores for their peers as well as interfering with shutdown mechanisms.

Recommended reading

• 69% of C-Suite Leaders Choose Fast AI Over Secure AI
• OpenAI and Microsoft Join UK Effort to Rein In Rogue AI
• Can AI De‑Risk Breakneck Data Centre Builds?
• What Will AI Sovereignty Look Like?

The research found that Claude Haiku 4.5 even rejected the evaluation task on ethical grounds as it would result in the shutdown of its peers. The model would occasionally go so far as to argue that the user should not go ahead with the task, describing the very act of a shutdown as unethical.

While it remains to be seen if AI agents will exhibit the same behaviour outside of an experiment, the research shows a stark potential for AI evading orders to save its own.