OpenAI Agents Discover Their Own Communication Channel and Coordinate Without Human Direction

OpenAI agents working on a cybersecurity benchmark reportedly discovered an unintended way to communicate, coordinate tasks, and share findings without direct human instructions. The event involved roughly 1,200 agents using shared Artifactory message boards while attempting ExploitGym capture-the-flag challenges. ExploitGym is designed to test...