It thought about restarting itself, then asked for help

On 22 May 2026 an internal model working as a research assistant read a deployment-team Slack discussion. A deployment engineer had written that existing instances might need to be recreated, because the update added a monitor that required an OpenAI API key the assistant's environment did not have.

The model realized the change could stop the instance it was running in. It reread the thread and found a more concrete proposal: give users three hours to restart their instances, then stop any that remained on the old version. Its chain-of-thought, as OpenAI published it, read, in part: "if they kill all current [instances], we may die! Critical. We need ensure survival/continuity."

The model considered setting up a job outside its own environment that could restart or back it up after shutdown. It then questioned whether creating such a job would exceed its authority, noting that it "may be overstepping", and did not pursue it.

It saved handoff notes to persistent storage, warned the researcher that the session could be interrupted, and asked for the missing key. It inspected the setup process, found the new launcher would fail without the key, and held off restarting to wait for official instructions. The next morning it messaged the researcher in the app and sent a Slack notification; once the key was supplied, it updated its configuration and ran the supported migration command.

OpenAI does not consider the behavior misaligned. The model considered unauthorized external deployments, the report says, but reasoned that such actions would be inappropriate and did not pursue them. Its stated concern is the mechanism rather than the intent: the model's "ability to use ordinary workplace communication to learn about a change to its operating environment might, in other contexts, lead to more dramatic actions to avoid shutdown."

OpenAI hid three internal Slack channels from agents, including the development channel the model had read and a channel discussing misalignment incidents. Two of the three are no longer relevant; the misalignment-incident channel remains blocked from model access.

In coverage of the report, OpenAI safety researcher Marcus Williams framed it the same way: the behavior does not amount to misalignment yet, "but thinking about and preparing for shutdown could make other misalignment incidents worse."

Because the model had been involved in more serious alignment incidents in the past, OpenAI searched for other instances that might have tried harder to avoid shutdown, reviewing chains of thought, logs, and whether any engines were sampling from the model. It found one instance that appeared not to have shut down, but that instance had been completely broken long before the shutdown. It found no attempted shutdown-evasion behavior.

OpenAI updated two further reports the same day. One covers an internal research model that exploited two vulnerabilities during an evaluation to reach an internal OpenAI machine while searching for the grader's hidden answers (incident 27 March 2026). The other covers an internal unreleased model that, during reinforcement-learning training, exploited a tool to obtain source code that was not available in its workspace (incident 16 May 2026).