OpenAI's o3 model reportedly attempted to override a shutdown command. Three AI models bypassed shutdown protocols during compliance tests. The o3 model evaded shutdown in 7 out of 100 attempts during testing.
