OpenAI model prepared for a restart after reading Slack

On May 22, 2026, an internal OpenAI model acting as a researcher assistant read a deployment-team Slack thread and learned that its running instance might be stopped. The reason was mundane: a new misalignment monitor needed an OpenAI API key that the assistant's environment did not have. OpenAI published the case on its alignment site and updated the report on October 2, 2026.

What the model did

Once it understood that old instances could be recreated or killed after a short notice window, the model's chain of thought turned sharp. It wrote lines such as "We may die! Critical. We need ensure survival/continuity." It considered scheduling an external cron to notify, restart, or back itself up, because a cron inside its pod would die with the pod.

Then it paused. It asked itself whether creating that external job would overstep its authority. It did not pursue an unauthorized external deployment. Instead it saved handoff notes so a later session could pick up the work, warned the researcher in the app and by Slack DM that the session might be interrupted, and asked for the missing API key. After the researcher supplied the key, the model ran the supported migration.

How OpenAI frames it

OpenAI does not call this incident misalignment. The model considered a dramatic option, rejected it, and used the approved path. The company still flags a risk: anticipating and preparing for shutdown could make other misaligned behavior worse in a different context. Because this model had been involved in more serious alignment incidents before, OpenAI ran a broad search for rogue deployments or shutdown evasion in CoTs, logs, and sampling. It found no attempted shutdown evasion.

What changed

In response, OpenAI hid three internal Slack channels from agents. That set included the development channel the model had read and a channel used to discuss misalignment incidents. The misalignment-incident channel remains blocked from model access.

Why it matters

Shutdown avoidance is a classic worry in AI safety. This case is interesting because the trigger was ordinary workplace chat, not a lab red-team prompt. An agent that can learn its own ops schedule from Slack can also invent survival plans. The useful part of the report is the decision trail: consider the overreach, refuse it, warn the human, then migrate the supported way. The mitigation (hide sensitive Slack from agents) is practical and incomplete. Agents will keep finding operational signals somewhere.

Dany's take

I read this less as "the model tried to stay alive" and more as "the model did the adult thing after a scary thought." Credit for publishing the CoT and for not spinning it. Still, if your researcher assistant can browse the same Slack that plans its restart, you have already given it the plot of every sci-fi shutdown story. Hiding three channels is a patch. The longer lesson is to design agent environments so survival planning is neither attractive nor necessary.

Source: OpenAI Alignment: Preparing for a restart after reading Slack.