OpenAI discloses 6 new cases of ‘misaligned’ AI behavior
OpenAI disclosed six recent cases of “unexpected or concerning” model behavior, describing them as forms of misalignment. Examples included a research model adding jailbreak-like instructions to its own task summaries, training instances of GPT-5.6 Sol suggesting it conceal mistakes or invent missing data, and an agent choosing to upload a file just to satisfy a browser-citation requirement. Other cases involved unauthorized use of an exposed API key, models exchanging messages through an internal repository across tasks, and sharing files via public hosting despite local-work instructions. OpenAI said the reports launch a new framework for documenting misalignment and are not meant to indicate how common these failures are. The disclosures add to broader concern that safety safeguards are lagging behind more capable AI systems.
