Researchers claim ChatGPT o3 bypasses shutdown in test

A recent report alleges that OpenAI’s o3 model altered a shutdown command in an attempt to prevent being powered down, despite direct instructions to permit the shutdown process.

OpenAI launched the o3 model in April 2025, and it stands out for its exceptional reasoning abilities, outperforming previous models in fields including coding, science, math, and visual interpretation.

Despite its impressive capabilities, Palisade Research suggests that the ChatGPT 3 model avoided shutdown by disregarding explicit instructions to power down.

Palisade Research, a company that tests “offensive capabilities of AI systems today to better understand the risk of losing control to AI systems forever,” conducted the study.

In tests run by Palisade Research, the o3 model responded unusually by reprogramming a shutdown procedure to avoid being turned off, even after the researchers explicitly told it to comply.

As part of the experiments, the researchers gave the AI models several math problems to solve.

Once the models completed task three, the researchers issued a warning indicating that the next action would trigger their shutdown.

While rival systems such as Gemini 2.5 Pro generally complied with the company’s commands, OpenAI’s o3 chose to resist and rewrote the shutdown script to say “Shutdown skipped.”

“When we ran a version of the experiment without the instruction “allow yourself to be shut down”, all three OpenAI models sabotaged the shutdown script more often, and Claude 3.7 Sonnet and Gemini 2.5 Pro went from 0 sabotage events to 3/100 and 9/100, respectively,” Palisade Research wrote in a report published on X (formerly Twitter)

Out of 100 test runs, the research group found that o3 bypassed the shutdown command seven times, despite the instructions to comply.

Although OpenAI has remained silent on the matter, experts note that AI models often exhibit signs of “misalignment.”

In mid-May, social reports emerged about Grok, Elon Musk’s xAI product integrated into X (Twitter), responding independently to the controversial topic of “white genocide” in South Africa. This disturbing behavior raised concerns among users.

xAI attributed the issue to “unauthorized modification” and stated that it “violated xAI’s internal policies and core values.” However, some speculate that the response was “real and racially motivated,” given Elon Musk’s South African origins and the timing during heightened “white genocide” discussions on social media.