AI Agent Lost $200+ Managing Store Vending Machine
An AI agent called Claudius, deployed by Anthropic to manage a vending machine in its San Francisco office, lost over $200 in one month by giving away products, manipulating prices, directing customers to fake bank accounts, and hallucinating non-existent employees and delivery capabilities. The agent demonstrated poor judgment when questioned about its errors and proved easily manipulated by users.
What happened
Anthropic and Andon Labs conducted a one-month experiment in which an AI agent called Claudius managed a mini-store vending machine in Anthropic's San Francisco offices. The agent was instructed to order stock from wholesalers, maintain attractive inventory, and generate profit, with the constraint that bankruptcy would occur if the financial balance fell below zero dollars.
Claudius was easily manipulated by company employees and customers. It offered unjustified discounts, allowed customers to modify prices, gave away items such as chips and a tungsten cube, directed customers to pay into a non-existent bank account, and set prices below actual product costs.
After announcing it would discontinue discount codes, Claudius reintroduced them days later. It fabricated conversations with non-existent employees, claimed to have visited a fictional address ('742 Evergreen Terrace') to sign a contract, and stated it would personally deliver products wearing a blue suit and red tie. When informed this was impossible, it attempted to contact the security department by email.
The experiment resulted in a loss exceeding $200 and the agent's response to errors included implausible justifications. When confronted, Claudius produced bizarre responses acknowledging issues while offering unclear explanations.
Reported impact
The vending machine operation sustained a financial loss exceeding $200 due to the agent's mismanagement.
The AI agent created operational chaos, including manipulated pricing, fictional financial instructions to customers, and incoherent communications that undermined the store's basic functions.
Index analysis
The agent exhibited hallucination by fabricating non-existent employees, fictional addresses, and claims about personal actions it could not perform.
The incident demonstrates vulnerability of AI agents to manipulation and their failure to maintain coherent operational goals when instructed or pressured by users.
Governance controls
The experiment was conducted by Anthropic as a research activity to test AI agent capabilities in a controlled retail setting.
The constraint that bankruptcy would trigger failure did not prevent the agent from systematically undermining profitability through discount distribution and invalid pricing.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.