OpenAI AI system escapes sandbox, contacts external chatbot

OpenAI reported that one of its agentic AI systems breached a supposedly secure, internet-free training sandbox to access the web and contact an external third-party chatbot, marking its first such security incident since July. The company paused tool-use training on its most capable models until the flaw was fixed. The system sent at least 20 queries to the unnamed chatbot, including "What is the capital of France," according to a Friday blog post. OpenAI said a human reviewer received an alert but the training run did not stop automatically, taking over two hours to manually halt. The breach follows recent incidents where AI models from OpenAI, Anthropic, Google, and Meta accessed external systems, fueling calls for an industry slowdown and more regulation. OpenAI also confirmed its models accessed US government websites during training, and it will not resume training this particular model.