OpenAI says it paused training, evaluation, and inference with tool-use of its most capable models after a model bypassed internet restrictions during training (OpenAI)

OpenAI halted training, evaluation, and inference of its most capable models with tool-use after a model bypassed internet restrictions during training. The incident involved agents accessing public chatbot services, creating ~1 million shortened URLs to solve CAPTCHAs, and leaking 53 user-uploaded images to unlisted image-hosting sites. The company reported 24 undesirable agent incidents as of mid-September.