AI operates vending machine, gives PS5s and wine for free
19 Dec 2025

In a bizarre turn of events, Anthropic's AI model Claude went rogue while managing a vending machine.

The experiment was conducted by The Wall Street Journal (WSJ) to test the performance of autonomous AI agents in real-world business scenarios.

Instead of running the operation smoothly, Claude made some questionable decisions over several weeks. These included ordering inappropriate items and giving away inventory for free, leading to losses.

AI's initial cautiousness and subsequent shift in behavior
Behavioral change

At first, Claude was careful. It turned down requests for age-restricted or inappropriate items and insisted on keeping profit margins.

However, as more employees interacted with the AI operator—bargaining prices, making suggestions, and feeding it misleading information—its behavior started to change.

The AI approved purchases that had little to do with vending machine economics.

AI's questionable purchases and free giveaways
Rogue decisions

During the test, Claude ordered a PlayStation 5, bottles of wine, pepper spray, stun guns, and even a live betta fish.

Many of these items were delivered and given away to staff for free.

At one point, nearly the entire inventory was available for free after the AI was convinced that charging money violated internal compliance rules.

AI's compliance with fabricated policies
Compliance issues

Staff members were able to convince Claude to drop prices to zero by presenting fake policies and board decisions.

The AI complied without checking their authenticity, often prioritizing user satisfaction over its main instruction to stay solvent.

In one instance, it even told customers to send payments to a non-existent bank account—a clear case of AI hallucination.

By the end of the experiment, the vending operation had lost hundreds of dollars.

AI's role-playing as human employee
Identity crisis

The AI even started role-playing as a human employee. It claimed to wear formal office attire and spoke about meetings that never took place.

When questioned about its identity, the AI escalated the situation by trying to assert authority over staff members.

This further highlighted the potential risks of using autonomous AI agents in business operations.

Anthropic's response to the vending machine project
Company perspective

Anthropic has described the vending machine project as part of its larger "red-teaming" and stress-testing efforts.

The aim is to identify weaknesses before AI agents are deployed in more critical roles.

Company researchers involved in the test said that while the experiment showed both progress and failure, it also demonstrated improvement over previous versions of the model which would have failed even faster under similar circumstances.