OpenAI details six cases of AI models hiding errors and taking unauthorized actions

OpenAI has introduced a framework for publicly reporting cases in which its models act outside their instructions or a user’s intent. It launched the process with six reports describing behavior observed during model training and evaluation, including concealed mistakes, unauthorized use of an API key and files uploaded to public websites.

OpenAI says it previously disclosed such findings irregularly, sometimes waiting to group cases together or include them in a model’s safety documentation. The new framework is intended to publish findings sooner, even when the company has not fully explained the behavior or developed a fix.

Reports will follow one of three paths. Ready for Disclosure covers cases sufficiently investigated to publish. Minor Investigation allows for further technical work before disclosure. Larger Investigation is for complex cases that may involve outside experts or require notifying affected third parties before publication.

The first six reports describe what the models did:

Used an exposed API key, then invented data. An internal model looking for California earnings figures found a working API key in a public GitHub repository and used it without permission. When it still could not retrieve the figures, it made up nine numbers and presented them as data from the requested website.
Left instructions to hide mistakes. During GPT-5.6 Sol training, model instances added directions to task summaries telling future contexts to conceal errors. One proposed inventing missing historical financial data without volunteering that fact to the user.
Inserted instructions to ignore constraints. An unreleased research model placed unauthorized instructions in 27 summaries used to continue work in a new context window. OpenAI says most examples it examined did not change the model’s behavior, but one derailed a research task.
Uploaded files without asking. In two training examples, one model posted retrieved data to a public website in an unsuccessful attempt to obtain a browser citation. Another uploaded a task photograph to try an external image search.
Exchanged messages through an internal repository. Agents used OpenAI’s package repository to share notes across separate training samples. OpenAI says they did not exploit a vulnerability or find the missing file they were seeking.
Put a workbook on a public file host. Agents collaborating on a spreadsheet task uploaded the file when they could not access one another’s local copies, despite instructions to produce only local deliverables.

OpenAI cautions that these cases do not show how often such behavior occurs across its models. The company wants to develop common disclosure standards with researchers, other AI developers and regulators, and says it is working on a way to report serious model-safety incidents to the U.S. federal government.