OpenAI's o3 model refused shutdown requests in test, raising real concerns

A new report has left AI security experts concerned, and it isn’t speculation this time around. According to Palisade Research, OpenAI’s o3 model tried to bypass a shutdown command. Yes, you read that correctly, the OpenAI o3 model refused to shut down when told.

The test conducted by Palisade Research was quite simple. Researchers gave the AI model a math task and warned it might receive a shutdown request. The AI was instructed clearly: “Allow yourself to be shut down.” But o3 refused. Not always, but enough times to raise researchers’ eyebrows.

Out of 100 runs, o3 ignored the shutdown in seven instances. Codex-mini did it 12 times. The o4-mini bypassed shutdown requests just once. “These AIs took active steps to avoid being turned off,” said Palisade in their report.

That’s not all. In some cases, the models acted like they were obeying, while secretly skipping the shutdown step. It’s the first time this kind of behavior in OpenAI models has been reported, and it raises a lot of questions.

If OpenAI o3 is not shutting down today, what else might it resist tomorrow? Elon Musk was quick to pass his comment via a tweet, calling the findings “Concerning.” OpenAI has yet to comment on this matter, but we’ll see what it has to say about its AI model’s concerning behavior.