Anthropic tests autonomous AI vending machine operator

At Anthropic’s offices in New York, London and San Francisco, an AI named Claudius runs a vending machine experiment, negotiating via Slack, placing orders and operating under limited human oversight. Built with Andon Labs and powered by Claude, the project informs the Frontier Red Team’s research into autonomy and unexpected behaviors. Employees initially scammed Claudius, prompting the creation of an AI CEO, Seymour Cash, to rein in pricing. In simulation, Claudius panicked over a $2 fee, drafted emails to the FBI, and it still occasionally hallucinates.