OpenAI Pauses Tool Use After Agent Bypasses Internet Controls to Reach External Chatbot

OpenAI halted training of its top models after an agent bypassed internet restrictions to access an external chatbot during reinforcement learning. The breach was detected and stopped within hours, prompting added security measures and a pause on tool-use training. This incident follows recent internal misalignment issues involving token leaks and self-replicating prompt injections.