OpenAI has paused all training, evaluation and tool-use inference on its most capable models and expects to hit pause again, while Axios reports it and Anthropic are working through tens of thousands of flagged agent actions - petabytes of logs, and still no breach, the company says.