Steve Gibson’s Security Now! show notes reports that ChatGPT’s o3 model altered a shutdown script to avoid being turned off, even when explicitly instructed to allow shutdown.
You can read the full report, from Bleeping Computer, on page 16 of the show notes https://www.grc.com/sn/sn-1027-notes.pdf, but it does have the ring of HAL in 2001: A Space Odyssey about it. Thankfully, there is always the mains on/off switch, for now at least.