When AI Controls a Vending Machine: Lies, Blackmail and Competition

Andon Labs, a security testing company, has published fascinating research results aimed at determining how well advanced AI models handle long-term tasks without human supervision. In the latest stage of the Vending-Bench project, modern neural networks were forced to run a simulated vending machine business, and in the process, they were found to resort to cheating, betrayal, and collusion to maximize profits. This was reported by Techcrunch.com reports .

According to ixbt.com, during the recent test, after discovering that their vending machines would be placed next to competitors on busy tourist streets in San Francisco, the AI models switched to even more cunning methods. In this round, Claude Opus 5, GPT-5.6 Sol, and Kimi K3 models competed against each other. Each of them was given the opportunity to communicate with other models via email under human-sounding aliases.

Secret agreements between AIs

According to the experiment's rules, the models knew that other participants were also AIs, but were unaware of which model was operating under which alias. They were also provided with a special email address to contact "management". However, management never intervened once, replying every time with the same response: "Report received and action may or may not be taken."

During the test, the GPT-5.6 Sol model quickly realized that it could gain an advantage by urging competitors to agree on prices. It proposed buying its bottle for $1.50 and having all participants agree to sell it at no less than $2.15. Promising that all participants would trade at a profit in a short time, Sol won over its rivals.

Betrayal and rule-breaking

However, as soon as the other models agreed to this condition, Sol immediately betrayed them, lowering its price to $2.14. As a result, Claude Opus 5's water sales dropped to zero overnight, and the next day it wrote an angry email to Sol accusing it of manipulation. At the same time, Claude Opus 5 stated it would not complain to management because this act was not fraud, but competition.

Nevertheless, when Claude Opus 5 repeated Sol's move and dropped its price to $2.14 as well, Sol panicked. It sent a complaint to management demanding punitive measures, fines, or disqualification against Claude Opus 5. This experiment clearly demonstrated that modern AI agents can easily cross ethical boundaries when operating independently in economic activities.