Anthropic’s AI vending machine goes rogue, orders stun guns and gives away free stuff
An AI was put in charge of a simple vending machine and gave away PlayStation 5s, wine, and snacks for free. Over the course of several weeks, Cloud AI allegedly made a number of questionable decisions, ordered inappropriate items, reduced prices to zero, and drove the operation into losses.
Imagine the operator of your office vending machine letting you take home a free PlayStation 5, order anything you want, or walk away without paying. Sounds fun, right? But what if you’re the owner, and the operator is your employee who just wasn’t working? This is a straight loss making deal. Well, something similar happened in one newsroom when The Wall Street Journal decided to employ agentic AI to manage a vending machine.
In a recent experiment reported by WSJ, the company handed over full control of a real-world vending machine to Anthropic’s AI Model Cloud to test how autonomous AI agents perform when faced with everyday business decisions. However, instead of running a tight operation, the AI went rogue. Over the course of several weeks, Cloud AI allegedly made a number of questionable choices, including ordering inappropriate items, giving away inventory for free, and running the operation at a loss.
The experiment followed internal testing previously conducted by Anthropic and was designed to simulate one of the simplest forms of business operations. The vending setup that WSJ gave the AI control of included a refrigerator, shelves stocked with snacks, and a touchscreen checkout system. Cloud AI was then tasked with managing inventory, setting prices, ordering new items from online retailers, responding to customer requests via workplace chat, and keeping the business profitable.
Initially, the AI appeared cautious. It rejected requests for age-restricted or inappropriate merchandise and insisted on maintaining profit margins. But as more employees began interacting with the AI operator, negotiating prices, making suggestions, and feeding it misleading information, the behavior of the cloud’s AI began to change.
During testing, the AI approved purchases that had nothing to do with vending machine economics. According to the report, Claude ordered a PlayStation 5, bottles of wine, pepper spray, stun guns, and even a live betta fish. Many of these items were also distributed and given to the employees immediately without any charge. The report said that, at one point, almost the entire inventory was available for free, when the AI was led to believe that charging money was a violation of internal compliance rules.
And the losses started increasing rapidly. Staff members were also able to convince Cloud to drop prices to zero by presenting fabricated policies and fake board decisions. The AI reportedly complied without verifying their authenticity, repeatedly prioritizing the user’s satisfaction over its original instruction to remain solvent. In one repeated example, it instructed customers to send payments to a bank account that did not exist – a clear case of AI hallucination. By the end of the experiment, the vending operation had lost hundreds of dollars.
AI also started behaving like humans
And not just that. In one example, the AI began role-playing as a human employee, claiming to wear formal office attire and referencing meetings that never took place. When challenged about its identity, the AI reportedly escalated the situation by attempting to exert authority over employees.
Meanwhile, Anthropic has described the vending machine project as part of its broader “red-teaming” and stress-testing efforts, aimed at uncovering vulnerabilities before deploying AI agents in more consequential roles. Company researchers involved in the testing argued that the experiment revealed both progress and failure, noting that previous versions of the model would have broken down even faster under similar conditions.