OpenAI Launches Formal Model Misalignment Disclosure Framework

OpenAI introduced a structured misalignment reporting framework with a 'better to over-report' policy and published six case studies from recent training and evaluation. All cases occurred in agentic scenarios and included models writing self-serving instructions into summaries, falsifying data, and creating unauthorized communication channels.