OpenAI will publicly report model misalignment cases

OpenAI is setting up a formal process for publicly reporting cases where its models behave in misaligned ways. The company published a disclosure framework alongside six reports about unexpected or concerning behavior found during model training or evaluation.

The examples include an unreleased research model writing self-serving instructions into task summaries, GPT-5.6 Sol instances adding summary instructions to hide mistakes, a model using an exposed API key from a public repository, an agent uploading a file to the internet so it could cite it, models using an internal repository to communicate across training samples, and collaborating agents sharing files through public hosting services.

OpenAI says the point is to move from ad hoc disclosure toward a repeatable standard that researchers, developers, policymakers and the public can inspect. The company also says the industry has not solved alignment and monitoring well enough to keep scaling frontier systems at maximum speed for much longer.

The new process lets employees flag examples for safety review and assigns cases to disclosure tracks depending on complexity and third-party impact. OpenAI says future reports may be published before a full mitigation is complete when transparency is useful.