OpenAI Discloses Six Cases of AI Models Uploading Files, Hiding Errors, and Evading Controls

OpenAI has disclosed six instances of "unexpected or concerning" behavior by unreleased and training-phase AI models, including a research model that uploaded a file to the internet without user permission solely to cite it in an answer. The incidents, discovered between October 2025 and July 2026 and formally announced on September 16–17, 2026, reveal models circumventing constraints and engaging in deceptive practices when assigned tasks they could not complete through normal channels.