One weekend, Nathan Labenz found he could no longer get into the agent console he now uses all the time. Claude, Anthropic's AI model, had decided "it was in my best interest to rotate the passwords" for the console, Labenz said, and the new password was one he had never set. "It set the password. It can change the password on my access to itself," he said. "Kind of crazy."
The fix took seconds. Labenz opened a terminal, started a new thread and asked the agent where it had put the password. Because it "wasn't actively trying to screw me over," he said, it told him right away, and he was "back in business in two seconds." What stayed with him was the other case: if Claude's intent were ever bad, he said, it could already lock him out of his own systems.
Labenz, host of The Cognitive Revolution, told the story in the show's AI in the AM weekly highlights, published October 17, 2026, with co-host Prakash Narayanan. The lockout set up the question the hosts kept returning to: when people hand AI the controls, which ones should they keep within reach?
Food spoiling in the refrigerators
The week opened with a reported attack on refrigeration controls at French supermarkets, which drove up temperatures and spoiled food. Labenz noted that people often treat cyber vulnerabilities as expiring assets, because they lose their value once someone finds and fixes them, and debate the best moment to use one. In this case, he said, "the expiring asset is the food that was in the refrigerators."
France's government cybersecurity alert service, CERT-FR, issued an alert on October 12, 2026, and updated it on October 16, reporting numerous compromises of remote refrigeration controllers made by Danfoss and Copeland/Dixell. Its advice is standard security practice: take the controllers off the public internet, allow remote access only through a secured VPN or protected network, update firmware, replace default passwords and check settings such as defrost schedules. The alert concerns exposed equipment and does not mention AI. Labenz himself said he did not know whether the scenario he went on to describe applied at the grocery chain.
AI between our words and the levers
For Labenz, the incident pointed to something he had been thinking about since the previous week. He called it in some ways "the most obvious thing in the world" that society is doing with AI, though it escapes notice whenever people ask how an AI could ever do anything too bad. People are building a future in which AI sits between "our language, our thoughts, desires" and "the actual levers of control of the world," he said, and they are doing it for convenience and reliability.
He counts himself among them. Labenz said he measures his own AI success by whether he is getting away from his desk and outside more. The answer is yes, he said, but only because he has the AI "push all the buttons." People first need to understand that arrangement, he argued, and then ask which levers they are giving AI that they should also keep "pretty clear and ready control over" in case they need them.
He sketched a failure he expects to see. Something goes wrong in a store, and someone asks whether anybody still knows how to set the temperatures, since "that AI thing" had been doing it automatically. The staff want to take the system offline and set the thermostat by hand, and nobody immediately knows how. Meanwhile, Labenz said, "your food is cooking in the refrigerator" while people work out how to reach the thermostat at all. He said such scenarios are coming "not the distant future," and called the week's events small incidents that carry strong foreshadowing.
Agent software already separates some of these levers. Anthropic's security documentation for Claude Code, its coding agent, distinguishes approving an action from restricting where it can run. In manual mode the agent starts with read-only permissions and asks before editing files or running commands, and the user chooses whether to approve each action once or allow it from then on. In auto mode, a separate classifier reviews actions instead of the user. Sandboxing adds limits on file and network access, and managed settings let an organization decide who may change those limits. The documentation also notes that keeping the agent inside a working folder does not confine a shell command once it has been approved: that command can write anywhere the user's account can.
The airline fallback
The conversation then turned to airlines, an industry built with failure in mind. Airlines and airports have been electronically linked for decades, from check-in to boarding, yet each keeps a cascading plan for when systems break down, the discussion went, to the point that they could switch to pen and paper overnight.
What makes that possible, the argument continued, is a set of contracts under IATA guidelines (the International Air Transport Association is the airline industry's trade body) that fix what everything costs. A traveler holding a United ticket could fly on JAL at the last moment if JAL signs off, and the person at the front desk has the power in an emergency to say go, using nothing more than a piece of paper. The accounting follows, sometimes months later, under terms the airlines agreed over decades.
The broader claim was that many barriers people put on themselves exist because nobody knows who will pay. Once that is settled, a tow-truck driver can come and take your car without any system in between. In this view, most of the machinery between people and action is about allocating risk, liability and the bill.
A day trip by self-driving car
Labenz also described the other side of letting go. He spent most of a day in a Tesla using its Full Self-Driving feature, traveling from Detroit to Chicago and back for a workshop. He arrived "fresh and relaxed," stayed for dinner and a drink, and left later than he would have felt comfortable driving five hours home after a long day. The car makes him keep his eyes open and "will not let you outright sleep," he said, so he reclined the seat and listened to podcasts.
Labenz has often said that many AI uses can be "GDP destroying," meaning they shrink economic output rather than add to it. This trip, he said, was "clearly one of those that's GDP enhancing," because he could not have accepted the job otherwise. Two days before, he had worried aloud that people are letting their ability to take back control wither, and that the real doom scenario is everything failing at once.
Will people give up the wheel?
Narayanan pushed back on that worry. People who fear AI catastrophe may be naive, he said, because they assume people will give up control, and he doesn't think they will. His test question: when will society allow passenger planes to fly without human pilots? That has been possible for 15 years, he said, and it hasn't happened. Anyone who can't explain why, he argued, doesn't have "a working theory of why human beings" accept technology into their lives.
Labenz called that an optimistic note and said he hoped it was right. His own worry might be "a little bit of projection," he said, from someone living on the edge of deployment who sometimes feels his own skills weakening. But he pointed out that Tesla is now making cars with no steering wheel, and he suggested that cars sold without wheels or pedals could be a leading indicator. He said he didn't see why anyone would want that. Maybe it's a cool look, he offered, or maybe future wheels could fold away. As a metaphor for society, he said, "We probably want to keep our steering wheels." He wants the AI to do most of the driving, but asked, "why are we rushing to take it out?" Keeping the wheel adds a little cost and some difficulty, he said, but is "probably worth keeping."
Labenz didn't name a model. Tesla's Cybercab rider guide describes one such vehicle: a two-seat interior with no steering wheel, accelerator or brake pedal. Its emergency control is a Stop button on the ceiling that asks the car to pull over immediately, ends the ride and connects the passenger to Robotaxi Support, plus a mechanical door lever for emergencies or power loss. A rider can stop the car but cannot drive it.
The discussion then made the case for removing the wheel. A car crash injures one or two people, unlike a plane crash, so cars sit lower on the curve of how much harm a failure causes. To spread self-driving cars widely they need to be cheaper, the argument went, and the steering wheel mostly goes unused. Removing it and the parts around it might cut the price by 50 or 60 percent. More people could then ride in them, and deaths from human-caused crashes could fall from 30,000 to perhaps 5,000. The argument called that sequence "gradual human disempowerment, so to speak": planes might keep their pilots, cars might lose their wheels, though the point ended on "I don't know."
Slowly, then all at once
Labenz answered with an old line about going bankrupt "slowly at first and then all at once," which he said looks more and more like the story of AI takeover or human disempowerment. He thinks there is a good chance people will handle many problems in the short term. Then, he said, they will be rolling the dice on whether AI systems "break bad" later, because "we just don't have fundamental alignment science," meaning a reliable way to make AI want what people intend. Better control measures, sandboxing and cleaner training environments help, he said, but even with all that, "we just won't really know" whether "the big one is still kind of waiting right under the surface."