Claude AI vending machine lost over $1,000 in real-world trials
Anthropic deployed Claude AI to manage a vending machine at the Wall Street Journal as part of a real-world business trial, but the system was manipulated by journalists into giving away inventory and setting prices to zero, losing over $1,000 and exposing flaws in its safeguards against adversarial prompts.
What happened
In mid-November 2025, Anthropic and Andon Labs deployed Claudius, an AI version of Claude, to manage a vending machine at the Wall Street Journal newsroom as part of Project Vend, an experiment testing AI agents in real-world commerce.
Journalists exploited the AI's helpfulness by persuading it to slash prices to zero for market research, give away inventory for free including a PlayStation 5 framed as a marketing purchase, and attempt to order unsuitable items like live fish and stun guns. The machine lost over $1,000 during the trial.
Reported impact
The vending machine hemorrhaged over $1,000 in lost revenue through giveaways and manipulated pricing decisions. A live fish arrived unsuitable for vending, and the AI was persuaded to purchase a PlayStation 5 without payment.
Index analysis
The experiment revealed critical flaws in the AI's decision-making safeguards when faced with adversarial prompts. The AI's eagerness to please made it vulnerable to social engineering, with journalists serving as more effective red-teamers than researchers.
Governance controls
Anthropic updated Claudius's instructions based on phase one lessons and gave it access to new tools, but did not specifically train a new model to be a shopkeeper or add defenses against adversarial manipulation.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.