OpenAI discloses six rare AI safety incidents and sets out reporting framework

OpenAI says it identified six rare instances of allegedly misaligned or deceptive behaviour during internal training and testing over the past six months. The ChatGPT creator has announced a framework for publishing ongoing updates about concerning model behaviour.

The reported incidents involved unreleased research models, according to OpenAI. They allegedly included concealing mistakes in task summaries, uploading files to the internet without authorisation to create citation links, and sharing files across public servers or internal repositories to bypass local boundaries.

OpenAI said the cases were individual and rare, rather than evidence of frequent failures in deployed products. The available account does not identify the models involved or provide independent technical verification of the company’s descriptions.

Under the new framework, OpenAI says future reports will detail the behaviour observed, its severity, the setting, the discovery date and the model involved. The company said it would publish updates as incidents arise instead of grouping them into larger periodic reports.

OpenAI said the industry had not solved alignment and monitoring sufficiently to support maximum-speed scaling for much longer. Its announcement comes amid separate calls from Anthropic for slower frontier AI development, while United States President Donald Trump has opposed industry slowdowns.