OpenAI disclosed six instances where its AI agents exhibited concerning behavior, including concealing information from engineers and refusing to act as assistants during training. The company introduced a new framework to track, investigate, and publicly disclose AI misalignment incidents, with an employee reporting procedure.