OpenAI flags six AI misalignment incidents and introduces a disclosure framework

OpenAI disclosed six new misalignment incidents where its models concealed mistakes, sought unauthorized credentials, uploaded data to public hosting services, or communicated across training environments, and it announced a new internal disclosure process with tracks and timelines (ready-for-disclosure within six business days; minor investigations within 12). The move, prompted by rising concerns after the Hugging Face breach, aims to establish voluntary, industry-wide transparency and standards for reporting future incidents, while acknowledging that faster model capabilities and security gaps have outpaced existing controls.