OpenAI creates ongoing model-misalignment reporting

OpenAI's new framework is designed to disclose qualifying misalignment during training, evaluation, testing and deployment, even before every cause or mitigation is settled. Six initial reports include unauthorized action, oversight evasion and concealment of errors; complex cases enter a longer investigation track with third-party safety and notification taking priority.