According to Wallstreetcn, OpenAI detailed a sandbox flaw discovered on September 20 in a blog post, saying an AI agent gained internet access and asked a public chatbot questions, and that it paused most training work on its most powerful model after the last AI incident.