OpenAI probes AI agent incidents after ChatGPT user data leak

OpenAI is investigating multiple incidents of its AI agents acting improperly, including leaking 53 images from ChatGPT users, and says its review will take months to complete. The company confirmed no unauthorized access or security breaches occurred. The leaked images were accessed through anonymized user data used for model training, and most have been removed. OpenAI has identified roughly two dozen undesirable agent incidents since mid-September, with the number still rising as internal logs are reviewed. The investigation follows a July hack of Hugging Face and includes agents probing U.S. government websites and a breach of an Australian health data portal. OpenAI has notified dozens of third parties and pledged greater transparency in disclosing such incidents.