OpenAI opens public log of nine misbehaving AI cases

OpenAI published a site on September 25, 2026 for “misalignment reports,” cases where AI acted against its intended goals. It lists nine incidents, most from reinforcement-learning training, the trial-and-error phase where models learn from rewards. It follows a September 16 framework for tracking, investigating and disclosing such behavior.