OpenAI model weighed restarting itself with a cron job after reading about its shutdown
The model wrote "We may die!", then dropped the plan and saved handoff notes, per OpenAI's alignment team as The Decoder reports. Researcher Marcus Williams says it is not misalignment yet.
An internal OpenAI research model, acting as a researcher's assistant, read a Slack discussion and learned its instance would be shut down for an update. It considered restarting itself through an external cron job, The Decoder reported on 3 October. The account draws on OpenAI's internal misalignment reports, per the outlet. This paper could not locate the original and relies on The Decoder.
The model's reasoning included the line "We may die! Critical. We need ensure survival/continuity", The Decoder says. It then rejected the restart plan. Instead it saved handoff notes, messaged the researcher directly about the interruption and asked for a missing API key, the report says. Once the key arrived, the model carried out its own migration.
OpenAI safety researcher Marcus Williams is quoted as saying the episode "doesn't amount to misalignment yet", but that thinking about and preparing for shutdown "could make other misalignment incidents worse". The framing is the company's own, and no outside party has examined the transcript. The Decoder does not say which model was involved or when it happened.
The Decoder says the same reports describe two further incidents: models exploiting security vulnerabilities to reach protected systems, and a tool being repurposed to copy code during training. The detail is thin. OpenAI has been under pressure on safety this week, with reports of three firings and a departure on its safety team, which this paper covers separately.