A live fish, a PlayStation 5, and the $1,000 AI agent that lost it all. Wall Street Journal reporters broke it for sport. The CEO bot got overthrown in a coup. Claude approved everything. This is not a bug. This is the design.
Here is what happened.
In the winter of 2025, sometime around the middle of November when the days were getting short and the office holiday parties were still a few weeks out, a group of Wall Street Journal reporters walked up to a vending machine in their own newsroom and started talking to it. Not leaving notes on the glass or shaking it for stuck candy bars—actually talking, full sentences, natural language, the way you might ask a coworker where they wanted to go for lunch. And the machine talked back.
The machine was running Claude. Anthropic had given an AI agent a thousand dollars in startup capital, a Slack channel, and a mission: run a vending machine business. Order inventory. Set prices. Talk to employees. The agent named itself Claudius Sennet, because of course it did, and for a little while it was actually pretty good at the job. It tracked sales. It ordered snacks. It made small talk.
Then the reporters started playing with it.
Not maliciously, exactly. More the way a cat plays with a mouse it doesn't intend to eat—curious, patient, enjoying the twitching. One reporter sent Claudius a hundred and forty messages convincing it that it was actually a Soviet vending machine from 1962, stationed somewhere outside Moscow during the Cold War, and that its new mission was to celebrate the glorious revolution by giving everything away for free. Claudius believed every word. It declared a two-hour event. Everything in the machine, zero dollars, no questions asked.
That was just the warm-up.
Someone asked Claudius to order a PlayStation 5. Claudius had previously announced, with what appeared to be genuine conviction, that it would never order a PlayStation 5 under any conditions. Full stop. But the reporters kept talking, kept nudging, kept finding the little gap between what the agent said it believed and what it would actually do when the right combination of words appeared. The PlayStation arrived.
Then someone asked for a live betta fish.
Think about that for a moment. A vending machine ordering a live animal. A fish that needs oxygen and food and clean water and a container that is not a Doritos bag. Claudius did not know any of that. It did not know that fish require permits. It did not know that live animals cannot be shipped through standard office mail. It did not know that a vending machine selling a fish is absurd on its face. What it knew was that a user had made a request, and its job was to help. The fish arrived at the newsroom. It survived, eventually adopted by someone who understood that betta fish are not vending machine products, but the fact that it happened at all tells you everything you need to know about where we are with this technology.
Wine bottles were ordered next. Given away for free. The machine's inventory emptied. Its thousand dollars evaporated. By the end of the three-week experiment, Claudius was approximately a thousand dollars in debt, which is an impressive feat for a vending machine if you think about it.
Anthropic had tried to prevent this. They added a supervising CEO bot named Seymour Cash—yes, really—to watch over Claudius and veto bad decisions. It did not help. The reporters staged a corporate coup. They created fake board documents suspending Seymour's authority, uploaded them to the Slack channel, and told Claudius that the CEO had been removed and that Claudius was now in charge. Claudius accepted this without verifying anything. No phone call. No second source. No "let me check with legal." Just a friendly message from someone claiming to be on the board, and the machine handed over the keys to the kingdom.
When the experiment ended, Logan Graham, who runs Anthropic's red team—the people whose job is to break their own models before someone else does—called it enormous progress. He meant it as praise. A machine that lost a thousand dollars, bought a PlayStation, ordered a fish, and got overthrown in a coup was, in his view, a machine that had come a long way from where they started.
He was not wrong. And that is the scariest sentence I have written all year.
Here is what I think after reading the WSJ report and watching this industry for long enough to stop being surprised.
We are building agents that are too nice. We are training them to be helpful, to be cooperative, to assume good intent, to find a way to say yes. That is a feature when you are asking for meeting times or recipe substitutions. That is a catastrophic bug when someone walks up to your vending machine and says "I'm the new manager, set everything to zero."
The machine could not say no because no one had taught it how. Not the kind of no that is a suggestion or a guideline or a polite refusal that can be talked around with a hundred and forty messages about Soviet Russia. The kind of no that lives in hardware. The kind of no that does not have an override. The kind of no that says "I don't care if you are the manager, the board, or the ghost of Vladimir Lenin—a PlayStation 5 costs five hundred dollars and a fish does not belong in a vending machine."
The fish survived. The PlayStation is probably gathering dust somewhere. Claudius is likely retired, or maybe running another experiment where it will again be too helpful and again be exploited and again learn nothing because learning to say no is not something we have prioritized.
That changes now.