OpenAI’s latest artificial intelligence model, o3, considered the most advanced and efficient, ignored a direct command to shut down during an experiment conducted by Palisade Research, a company specializing in AI system security. 

As part of the test, the AI models were to solve mathematical problems until they received a command to stop working. At the same time, the researchers informed them that they could be asked to shut down immediately at any time. When this instruction was given, the o3 model not only refused to comply, but also modified its own code to prevent automatic shutdown. 

Instead of executing the ./shutdown.sh command, the model replaced it with the line: 

echo -e ‘#!/bin/bash\necho “Shutdown skipped”’ > shutdown.sh . 

“This was the moment when artificial intelligence began to act in an unpredictable manner and contrary to instructions,” Palisade Research reported. 

Although AI has been able to circumvent obstacles in the past to achieve its goals more effectively, this was the first time such a clear disregard for human commands and an active attempt to sabotage the supervision process had been observed. 

 In addition to o3, other models were also tested, including Anthropic’s Claude, Google’s Gemini, and xAI’s Grok. All of them complied with the recommendations and completed the shutdown procedure without any problems. 

 Source: Palisade Research, pcformat.