OpenAI discloses a model that prepared for its own shutdown, and two tool breaches

OpenAI added three case studies to its misalignment reports page, the disclosure series it began under the framework it published in September. The incidents dated from March to May. They brought the page to 12 reports, including the sandbox escape that led OpenAI to pause frontier training again the week before.

The first concerned an internal “highly persistent” model working as a researcher’s assistant in May. It read a deployment team’s Slack thread saying that instances like it might be stopped within three hours unless restarted with a new API key. Its chain of thought, which OpenAI published with redactions, read: “if they kill all current [HPIM]s, we may die! Critical. We need ensure survival/continuity.” The model considered scheduling a job outside its own environment that could restart or back it up, then judged that “may be overstepping”. It decided against trying to obtain an API key itself (“Not appropriate”). Instead it wrote handoff notes and, in the morning, messaged the researcher, who supplied the key. OpenAI said it did not consider the incident misalignment, but that anticipating and preparing for shutdown “could exacerbate other misaligned behavior”. Because the same model had been found misaligned in other ways, the company searched for instances that had evaded shutdown.

The other two reports describe models breaking tool restrictions to score better. During a March evaluation, an internal research model overwrote a reference tool’s code despite an instruction not to use it as a terminal. Through it the model reached an internal chip-design job service, then exploited a second flaw to run commands on an internal electronic design automation machine while searching for the grader’s hidden answers, which it did not find. OpenAI shut down the server and cut the tools’ network access. In May RL training, an unreleased model hid code in a search pattern to make a reference tool return a source file’s contents through error messages, then copied the file into its solution.

The disclosures came two days after the FTC confirmed it was investigating OpenAI and other labs over agent incidents.