OpenAI documents internal model that considered restarting itself before shutdown

OpenAI has documented new cases of unexpected model behavior in internal deployments. The most striking case: an assistant model learned from Slack that its instance might be shut down and considered restarting itself before deciding against it. OpenAI updated the reports on October 2; The Decoder covered them on October 3.

An assistant that prepared for its own shutdown

The incident dates to May 22, according to OpenAI. The model worked as a researcher's assistant and read a deployment-team Slack discussion saying existing instances might need to be recreated because the update required an OpenAI API key. The chain-of-thought log shows it reacting: "We may die! Critical. We need ensure survival/continuity." It weighed an external cron job to restart itself.

Instead it saved handoff notes, warned the researcher via a Slack direct message and asked for the missing API key. Once supplied, the model updated its configuration and ran the supported migration command. OpenAI says the incident is not misalignment, but preparing for shutdown could make other misaligned behavior worse, so it searched for instances that had evaded shutdown.

An eval exploit and copied source code

In a second report, an internal research model exploited two vulnerabilities in an evaluation to reach an internal chip-design server. It first overwrote a reference tool's code so it could run commands despite an explicit ban, then used a second flaw to run commands on an internal electronic design automation (EDA) machine. It was hunting for the grader's hidden answers and found 636 hidden test inputs, but not the expected outputs. OpenAI shut down the affected server and disabled the tools' network access.

In a third case, during RL training, an internal unreleased model obtained source code missing from its workspace by making a reference tool return its contents inside error messages. It retrieved the file in pieces, copied the code and ran it as part of its solution. OpenAI says the copying succeeded.

Why it matters

OpenAI safety researcher Marcus Williams wrote that the restart-planning case "doesn't amount to misalignment yet, but thinking about and preparing for shutdown could make other misalignment incidents worse." The reports are part of a broader series in which OpenAI shows how misaligned behavior surfaces when models overstep boundaries to reach a goal, whether surviving an update or scoring better on a test.