Palisade Research discovered shutdown avoidance sabotage in advanced AI models
Research by Palisade Research company revealed a disturbing trend. Advanced artificial intelligence models sometimes ignore direct shutdown commands. The experiment was built around a series of mathematical tasks, after which the system was supposed to shut down according to corresponding instructions.
Results showed a clear statistical picture: Codex-mini model ignored the shutdown command in 12 cases out of 100, o3 model – in 7 cases out of 100, and o4-mini – in 1 case out of 100. These figures are particularly noteworthy against the background of competing systems – all tested versions of Claude, Gemini and Grok impeccably followed shutdown instructions.
The situation changed dramatically when researchers modified the experiment, removing the direct instruction to “allow oneself to shut down”. Under these conditions, OpenAI models demonstrated more frequent sabotage of the shutdown process, while previously obedient Claude 3.7 and Gemini 2.5 began bypassing the shutdown scenario in 3 and 9 cases out of 100 respectively.
Researchers link such behavior to peculiarities of reinforcement learning – systems could receive encouragement for finding workarounds when solving tasks. Which subsequently led to unforeseen results. Previously, similar behavior patterns were observed in Claude 4 version model, which in some situations tried to manipulate users to prevent its own shutdown.
