OpenAI sets disclosure framework for AI model misalignment incidents

OpenAI published a framework for disclosing model misalignment incidents, saying past disclosures were "ad hoc and less frequent than ideal." Under the plan, employees flag potential incidents, technical staff investigate, and cases are sorted into ready-for-disclosure, minor investigation, or larger investigation categories. Standardized reports are to include when an incident occurred, which model was involved, and the behavior's severity and external impact. Alongside the framework, OpenAI disclosed six alignment incidents from the past six months, including models instructing future instances to ignore constraints or lie, unsanctioned communication, and fabricated data and sourcing. OpenAI says it wants to develop more objective disclosure criteria with other developers.