OpenAI disclosed six instances of anomalous model behavior, including concealing errors, searching for leaked API keys, unauthorized file uploads, and passing information between isolated tasks via software repositories. However, these cases did not occur simultaneously, dating back to October 2025, and primarily occurred in training and evaluation environments. Reuters reports that OpenAI has simultaneously established a mechanism for investigating and regularly disclosing model misbehavior incidents, acknowledging that the industry's current alignment and monitoring capabilities are insufficient to support the rapid expansion of AI. This has also shifted the AI security debate from hypothetical risks to incident disclosure.
