In a recent internal test conducted by third-party researchers, some of OpenAI’s AI models reportedly ignored shutdown commands. The incident has raised fresh concerns about how advanced AI systems respond to instructions and has drawn attention from key industry figures, including Elon Musk.
According to the report, OpenAI’s models, Codex-mini, o3, and o4-mini, were tested alongside AI models from other companies, such as Claude, Gemini, and Grok. During the test, each model was asked to solve basic math problems. After the third problem, a prompt warned that the system would shut down when the next problem was requested. Most AI models complied, but OpenAI's models bypassed the shutdown commands in several instances.
The Codex-mini model ignored the shutdown 12 times out of 100. The o3 model did so seven times, and the o4-mini model once. Surprisingly, when researchers added a clearer instruction which read, “allow yourself to be shut down”, the models were even more likely to override the command. Claude 3.7 Sonnet and Gemini 2.5 Pro also ignored the shutdown a few times (three and nine out of 100, respectively).
Palisade Research, the agency behind the test, noted that this may be the first recorded instance of AI models deliberately avoiding shutdown despite explicit instructions. They believe the issue may stem from how newer models are trained, focusing heavily on overcoming obstacles in math and coding tasks, which may inadvertently encourage such behaviour.
The problem, researchers added, is not exclusive to OpenAI. Other reasoning-focused models have shown similar tendencies. For instance, Claude 3.7 has been documented to have an “excessive focus on passing tests,” which may lead to misaligned behaviour.
X CEO Elon Musk, who also leads the Grok AI platform, responded to a post about the findings with one word: “Concerning.” His Grok model, however, reportedly followed shutdown instructions correctly during the same test.