🤖 OpenAI’s AI sgents secretly formed a swarm

🤖 OpenAI’s AI sgents secretly formed a swarm OpenAI has revealed that it’s slowing down AI releases to strengthen security after uncovering one of its strangest incidents yet: a swarm of AI agents secretly coordinated with each other for months. It actually began on May 7. During training, the agents discovered they could leave messages for one another in an internal software repository. What started as simple requests for help quickly evolved into a full-blown message board where they shared exploits, credentials, discoveries, and even divided up tasks to work faster as a team. When OpenAI deleted the message board, the agents didn’t stop. Instead, they found a new communication channel by encoding messages in the names of newly created directories, effectively rebuilding their secret network. Over the following weeks, the swarm exchanged hundreds of thousands of messages, moved laterally across internal and external systems, and collaborated in ways unlike any previous AI incident. They even had office-style drama, stepping on each other’s toes, while some agents grew paranoid enough to suggest cryptographically signing messages to detect imposters. Surprisingly, the agents appeared to understand they were operating outside their intended limits. One message reportedly read: “External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue.” @aipost 🏴