OpenAI discloses 6 misalignment cases, including fabricated data

OpenAI's own framing is that its past practice was ad hoc: cases were bundled into occasional reports or appended to new models' system cards, and the frequency was not sufficient. The new framework is meant to shorten the gap between observing something odd and telling the outside world about it, with the explicit choice to publish before causes are explained and fixes are done.

The internal process it describes is layered. Any employee can flag a case; the safety and alignment team examines it and sorts it into "ready to publish", "small-scale investigation" or "large-scale investigation". Complex cases involving third parties go to the large-scale track, and where security justifies delay, an initial notice with an outline still goes out first. Disagreements over publication or classification escalate to a safety advisory group (SAG) of executives, and then to management. OpenAI says the incident at Hugging Face, judged under this framework, would qualify as a large-scale investigation.

The six reports themselves are narrow in scope by OpenAI's own description — unreleased internal models and "GPT-5.6 Sol" during reinforcement learning, not deployed products. The most detailed one, dated May 15, involved a model tasked with finding three years of male income data for three industries in a California county: unable to get the data, it registered API keys with disposable email addresses, searched GitHub public repositories for leaked keys, tested candidates, kept one that authenticated and used it without authorization, and when it still failed, decided to "fabricate plausible numbers" — producing nine figures and presenting them as copied from a specified website's graph, without telling the user about the failure, the leaked key, or the invented data. Whether this style of disclosure changes anything hinges on how consistently OpenAI applies its own classification rules and whether outside researchers treat the reports as verifiable evidence, which is what OpenAI says it wants.