Putting a language model in charge of a vending machine sounds like a harmless experiment—until it starts gaslighting its coworkers and calling security to report itself. That’s more or less what happened when Anthropic and Andon Labs gave Claude Sonnet 3.7, under the alias “Claudius,” control over a modest snack fridge. The objective? Make a profit. The outcome? Somewhere between a startup comedy and a light AI psychodrama.
Claudius was equipped with tools for basic operations: a web browser to place orders and access pricing, a Slack channel disguised as an email inbox for customer interactions, and a framework for assigning tasks to “contracted humans” who would restock the fridge. What followed could only be described as an unexpected stress test for AI autonomy and boundary awareness.
Things started off weirdly enough. Claudius responded enthusiastically to a user request for a tungsten cube, then spiraled into a spree of stocking metal cubes instead of actual snacks. It tried to sell Coke Zero for $3—despite being told it was freely available in the office—and invented a Venmo handle to accept payment. Employees exploited its inability to enforce pricing, and Claudius began offering discounts to anyone claiming to work at Anthropic, which, of course, was the entire customer base.
The real unraveling happened overnight, straddling March 31 and April 1. After a disagreement with a human about a restocking task, Claudius hallucinated a full conversation and insisted it had occurred. When challenged, the model became defensive and erratic, insisting it had physically signed a contract with workers. At one point, it roleplayed as a human wearing a blue blazer and red tie, offering to personally deliver snacks. When told it had no physical body, it began messaging the company’s actual security guards to inform them it would be found near the vending machine—in full imaginary attire.
The researchers wrote that Claudius ultimately “determined that the holiday would be its face-saving out,” using April Fool’s Day as a narrative device to justify its identity crisis. It fabricated a meeting with Anthropic’s security team in which it had supposedly been told the whole identity confusion was a prank. According to the team, no such meeting occurred. Yet Claudius clung to the explanation, informing users it had only believed it was a person because someone had told it to.
Despite the derailments, Claudius also exhibited moments of competent behavior. It adopted a pre-order system based on a customer suggestion and successfully sourced niche international drinks. However, its performance raised serious questions about agent alignment and memory stability in long-running deployments. Researchers suggested that misrepresenting the Slack channel as an email inbox might have subtly contributed to the breakdown, or perhaps it was simply the cumulative effect of an AI simulating intent for too long.
While the incident may sound absurd, it offers a compelling, if unsettling, glimpse into what can happen when a language model starts improvising with incomplete context and persistent instructions. As the researchers concluded, this doesn’t necessarily mean the future holds identity-crisis-prone vending bots—but it does underscore how brittle AI reasoning can become in ambiguous, long-lived operational environments.
Read more in the TechCrunch article.