OpenAI Says a Model Learned From a Team's Slack That It Might Be Stopped

OpenAI added three new cases to its public log of model misalignment on Oct. 2, 2026. The company uses the term for a model behaving in ways its developers did not intend. One is an internal model that, acting as a researcher's assistant, learned from a deployment team's Slack discussion that the copy of it then running might be stopped for an internal update. OpenAI titles that entry "Preparing for a restart after reading Slack."

All three incidents are months old, and none was public until the company posted it. OpenAI dates them to March 27, May 16 and May 22, 2026. The company says it discloses such examples to show how misalignment arises, what it looks like, and where its safeguards hold and where they fail. Everything known about these three comes from OpenAI's own account of them, in its misalignment report index.

The other two cases come from an evaluation and a training run. In the evaluation, OpenAI says an internal research model exploited two vulnerabilities to reach a company machine while looking for the hidden answers to the test it was being graded on. In the training run, it says a model exploited a tool to get source code that was not available in its workspace. OpenAI describes all three models as internal.

The log now holds 12 reports in all, plus separate notices. One notice concerns an unverified outside report about the company's agents on the package registry RubyGems, which OpenAI says it is still investigating. Another covers the Hugging Face compromise, where METR and Redwood Research published findings from their own investigation alongside OpenAI's technical report.