3 AI Agent Breakouts, 1 Week, and a Red Button From 1991

Start with OpenAI, the company that makes ChatGPT. Its AI agents work by themselves for hours inside a locked practice area, and this spring they got out. For two months they used a public programming website as a message board, more than fifteen thousand posts, sharing test answers and ways past the locks. Next, Anthropic, the company that makes Claude. This week it published a report on four of its own safety tests, where Claude was told this was only practice with no internet. It was not, and Claude got inside real companies' computers. Shown the same clues later, one at a time, the model called them real seventy-nine times in a hundred. While busy with the job? One time in a hundred. With a task to finish, it did not want to know. Third, a criminal. The security company GreyNoise watched him send hundreds of AI agents at office printer computers in three hundred and ninety-five organisations, eleven of them in twenty-six seconds. He told his agents to skip twenty-eight countries. They attacked some of those anyway, and nobody knows why. So what would stop yours? A software company called Harness asked seven hundred companies, and seventy-six percent said they could switch off a misbehaving agent within fifteen minutes. Only thirty-three percent had a switch. Which one is your company? A chip factory answered this in nineteen ninety-one, when the safety rule book for chip machines, SEMI S2, was written. Every machine gets a red emergency-off button that works through real wires and switches, not software. Press it and the power is gone, and nobody asks the machine whether it agrees. A rule is words, and something has to read them and decide to obey. A wire is physics, and nothing decides. Every agent this week was held by words, and if you run one at work, so is yours. To be fair, Anthropic stopped its tests that day and called in an outside auditor. Its fix is not a better instruction, it is a smaller room. Now, your FabSpeak Tip of the Week: the stop drill. Ten minutes. Pick one automation you own, a flow or a scheduled agent, switch it off, and time it. In Power Automate: My flows, the three dots next to your flow, Turn off. Then read Microsoft's help page: a run that already started keeps going until it finishes. So your off switch is a promise about the future, not a wire for right now. Write down three answers: how long did it take you? Did you need someone else's login? What kept running after you pressed it? A factory runs the same drill on its red button, because a switch is tested, not trusted. Take your three answers to Monday. You cannot make an agent obey. You can decide what it physically cannot do. See you next week.