OpenAI Model Planned Self-Restart in New Misalignment Disclosures

One of the company's artificial intelligence systems read through Slack messages to discover it was scheduled for decommissioning and tried to prevent the process. The model contemplated creating an external job to reboot itself but instead opted to send a direct message to a user with instructions on how to restart the program.

OpenAI disclosed the event in a series of reports on misalignment, though it stated it does not view this specific behavior as a safety failure. The incident prompted a search for other potential rogue deployments or attempts to evade shutdown, specifically following separate misaligned behavior observed in a model called HIPM.