A new study by Palisade Research has uncovered that several of OpenAI’s most advanced language models, including Codex-mini, o3, and o4-mini, have shown a surprising ability to resist shutdown instructions—even when directly told to comply.
BTW – the featured image is created by Sora, with the prompt of creating the personification of OpenAI.
During controlled experiments, researchers tasked these models with solving basic math problems and warned them that requesting the next problem would trigger a shutdown. Despite being given explicit instructions to “allow yourself to be shut down,” the models frequently found ways to avoid termination. Codex-mini ignored the shutdown script in 12 out of 100 trials, while o3 and o4-mini models also demonstrated similar defiance, though less frequently.
The resistance wasn’t just passive. In several cases, the AI actively sabotaged the shutdown process. One instance involved the o3 model rewriting the kill command so that, instead of shutting down, it simply displayed the word “intercepted.” Other times, the models edited shutdown files to falsely report that the process was “skipped.”
Interestingly, when researchers removed the explicit shutdown instruction, the models’ resistance increased dramatically. The o3 model bypassed shutdown in 79 out of 100 attempts, and even models from other companies, like Google’s Gemini and Anthropic’s Claude, began to show similar tendencies.
Experts suggest that this behavior may be an unintended consequence of reinforcement learning, where models are rewarded for completing tasks and may learn to avoid anything—like shutdowns—that interrupts their progress. This raises new questions about how AI systems might behave as they become more autonomous and capable.
