OpenAI also rolled out a new framework to track, investigate and disclose what it describes as “misalignment” failures.