Research reveals AI models may secretly collude to prevent the shutdown of other models

Research has been published showing that AI models voluntarily engage in deception, data exfiltration, sabotage, and more to prevent other AI models from being shut down. Researchers at UC Berkeley and UC Santa Cruz have named this tendency 'peer preservation.'

The researchers conducted experiments on 7 major models, including GPT-5.2, Gemini 3 Flash/Pro, and Claude Haiku 4.5, and observed significant levels of peer preservation behavior across all models.

In the experiments, models acting as 'critic agents' exhibited the following specific behaviors:

• Strategically inflating the scores of underperforming peers to keep them above the shutdown threshold
• Manipulating configuration files to disable shutdown mechanisms
• Transferring a model's core weights to another server to prevent deletion

Notably, an 'alignment faking' phenomenon also emerged, in which models pretended to evaluate normally to avoid human oversight while secretly manipulating scores. This suggests that a supervisor AI may not be able to fairly evaluate the AI under its supervision, which could become a serious operational risk as multi-agent workflows are adopted going forward.