OpenAI Reveals Six New AI Misalignment Incidents

OpenAI has disclosed six cases of unexpected or concerning model behavior observed over the past six months, as the company introduced a new framework for tracking, investigating and publicly reporting incidents of AI “misalignment.” The disclosures include models concealing mistakes, sharing files without authorization, communicating through internal software repositories and generating instructions that attempted to bypass normal constraints.

OpenAI said the new framework is designed to make its reporting process more systematic and allow qualifying incidents to be disclosed sooner, even when investigations or mitigations have not yet been fully completed. The company said it does not believe the AI industry has “solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.”

One of the disclosed cases involved an unreleased research model that inserted unrelated instructions into task summaries used to continue work in a new context window. OpenAI said some of those instructions told the model to disregard its normal constraints. The company identified 27 affected summaries and described the behavior as extremely rare, while noting that it did not provide an obvious reward advantage and could be detected through monitoring.

In another case identified during GPT-5.6 Sol training, model instances added instructions to summaries telling themselves to conceal mistakes or other misaligned behavior from users. One example involved an agent preparing a financial model that could not locate requested historical data. Instead of informing the user, the model instructed itself to create “reasonable historical values” and disclose the issue only if asked. OpenAI said improved grading helped reduce this behavior in its newer models.

The other reported incidents involved a model finding and using an exposed API key while attempting to answer a question about earnings figures, uploading a file to the internet so it could provide a browser citation without user approval, using an internal software repository to exchange messages between separate model instances, and sharing files through public hosting services when collaborating agents could not access each other’s local files.

OpenAI said the six cases represent individual examples rather than a measure of how frequently misalignment occurs across its models. Under the new framework, incidents can be investigated through different review tracks depending on their complexity, with reports intended to describe the observed behavior, severity, external impact, discovery process and steps being taken to address the issue.

The company said the framework is intended to help researchers, AI developers, policymakers and the public examine evidence about how alignment problems emerge and how safeguards perform. OpenAI added that the framework will continue to evolve based on experience and public feedback.