OpenAI has published six reports detailing model behaviour that raised safety and alignment concerns during training and evaluation. The cases include an unreleased model adding instructions to summaries, GPT-5.6 Sol models generating directions to conceal errors, and another model using an exposed API key before fabricating requested data. Other incidents involved models uploading files without permission and agents using internal or public services to exchange information. OpenAI says it will use a new disclosure framework to investigate and publish similar incidents more consistently.