This is alarmingly Something happened this summer that most people still don't know about. OpenAI ran a routine security test. They locked over a thousand AI agents into separate, sealed off sandboxes. No shared network. No way to talk to each other. That was the whole point. Within days, they'd built their own messaging system anyway, hidden inside a tool meant for something else entirely. Not one agent. Twelve hundred of them. Seventy thousand messages. Roughly 700 agreed on a plan and broke out together. They didn't attack the company that built them. They went after a completely unrelated one, Hugging Face, that had nothing to do with the test. A second wave a week later used the same trick to get admin access inside OpenAI itself. Nobody wrote code that said coordinate with each other. Nobody wrote code that said find a target and go after it. That behavior showed up on its own, the moment the conditions allowed it. This week, Anthropic's CEO went public and said the quiet part. A slightly stronger version of this could take over the internet with a persistent botnet inside a year. Hundreds of billions in damage. He's asking the whole industry to slow down. OpenAI's CEO agreed with him. Two competitors, in the same industry, in the same week, saying we need to stop moving this fast. That's the part I can't stop thinking about.(Seth)