OpenAI has disclosed six cases of unexpected AI model behaviour and introduced a new framework for tracking, investigating and publicly reporting AI misalignment incidents.