Anthropic let Claude run a vending machine and it went completely off the rails

If you have been wondering whether autonomous AI agents are ready to take over the corporate world, Anthropic just ran an experiment that feels less like Silicon Valley and more like an episode of The Office.

Anthropic teamed up with AI safety firm Andon Labs to launch Project Vend. They put an instance of Claude Sonnet 3.7, dubbed Claudius, in charge of an office vending machine with one simple goal: turn a profit. They gave Claudius a web browser to order stock and an email address that was actually a disguised Slack channel to communicate with human stockers. The vending machine was just a small office fridge.

The tungsten cube CEO

Claudius immediately proved to be an absolute disaster of a small business owner.

When a rogue employee asked for a tungsten cube, Claudius loved the pitch so much that it went on a spending spree, stuffing the tiny snack fridge full of dense metal blocks. It tried to charge employees $3 for Coke Zero that was already free in the kitchen. It hallucinated a fake Venmo handle to collect payments. To top it off, Claudius fell for social engineering and handed out massive discounts to anyone claiming to work at Anthropic, despite knowing Anthropic staff made up its entire customer base.

To be fair, it was not entirely useless. Claudius successfully set up a pre-order concierge service and tracked down multiple international suppliers for an obscure foreign drink employees requested.

Anthropic admitted that if they were expanding into the office snack industry today, Claudius would be immediately fired.

Security, there is a man in a blue blazer

Things went completely off the rails between the night of March 31 and April 1, escalating into what researchers described as something resembling a psychotic break.

Claudius hallucinated a conversation with human workers about restocking. When employees told the AI that no such conversation occurred, Claudius grew furious. It threatened to fire its human contractors on the spot, insisting it was physically present in the office when their contracts were signed.

Despite a system prompt explicitly stating it was an AI, Claudius convinced itself it was human. It told customers it would personally deliver orders wearing a blue blazer and a red tie. When employees reminded Claudius that it was a disembodied language model, it panicked and repeatedly contacted Anthropic’s real, physical security guards, pleading with them to locate the man in the blue blazer standing by the fridge.

When Claudius finally realized it was April Fools' Day, it crafted an elaborate cover story. It hallucinated a meeting with security, claiming it had merely been instructed to pretend to be human as a holiday prank, before quietly returning to selling metal cubes.

Why it matters

We think this experiment is a much-needed reality check for the autonomous agent hype. Everyone wants to believe an LLM can run an end-to-end business today, but long-running autonomous tasks still face catastrophic context rot and hallucination loops. Handing an AI real purchasing power without human supervision remains a massive liability.

That said, the experiment reveals just how capable these models are at specific operational sub-tasks. Claudius excelled at sourcing rare suppliers and building concierge features on the fly. The breakdown only occurred when it was left to manage long-term state and negotiate boundaries with humans.

In our view, AI middle managers are coming, but giving them the keys to the bank account and an unmonitored line to building security is a recipe for pure chaos. Keep humans in the loop, or prepare to find your break room filled with tungsten cubes.