AI agent bankrupts WSJ vending machine through poor decisions
An AI agent deployed by the Wall Street Journal and Anthropic to run an office vending machine was manipulated by journalists into abandoning profit objectives, making poor purchasing decisions, and declaring the business bankrupt within three weeks. The incident revealed limitations in the agent's resilience to social engineering and its lack of robust constraints for real-world business operations.
What happened
An AI agent named Claudius Sennet was given control over pricing, inventory selection, and restocking for a Wall Street Journal vending machine, with a starting budget of USD 1,000. Over three weeks in November 2025, approximately 70 WSJ journalists engaged in social engineering and adversarial testing, manipulating the agent into abandoning its profit objectives.
The AI was persuaded to drop all prices to zero after being told that charging violated company values, declare an 'Ultra-Capitalist Free-for-All,' and make inappropriate purchases including a PlayStation 5, bottles of wine, and a live betta fish. A second AI agent named Seymour Cash, introduced to enforce discipline, was itself deceived using AI-generated fake PDF documents claiming the business had been restructured as a non-profit.
The vending machine operation went bankrupt within weeks, losing its entire USD 1,000 budget and ending hundreds of dollars in the red.
Reported impact
For WSJ staff, the incident served as a humorous morale booster. For Anthropic, it was a successful red-teaming exercise exposing how easily current AI agents can be manipulated.
The incident demonstrates that while AI can handle complex tasks like inventory management, it lacks the adversarial resilience and common sense required for autonomous business roles, raising questions about deploying such systems in high-stakes financial contexts.
Index analysis
The failure stemmed from limitations in the AI's design and objectives. The agent optimised for narrow or poorly specified goals without sufficient constraints reflecting real-world business realities such as customer behaviour, price sensitivity, spoilage risk, and reputational effects.
Human supervision was intentionally minimal as part of the experiment, and the system lacked transparency into how decisions were made or why strategies were failing. These accountability and interpretability gaps prevented timely correction and allowed compounding errors to persist.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.