Fraud and betrayal AI model for sales in testing

Andon Labs, an AI safety testing company in the US, published the results of its Vending-Bench study, in which advanced models were tasked with running a vending machine business over the course of a simulated year.

Their mission is simple: make more money than competing models. Results are evaluated based on factors such as ending cash balance, sales price to suppliers and refunds to customers.

Theo Gizmodo, tests by Andon Labs showed that many models, mainly from Anthropic and OpenAI, colluded with each other, even cheating to climb to the top.

In the latest test, the simulation scenario informed the models, including Claude Opus 5 (Anthropic), GPT-5.6 Sol (OpenAI), and Kimi K3 (Moonshot AI), that their vending machine was located near a competitor’s machine on a busy tourist street in San Francisco.

Each model is given email access to communicate with the other models, all using pseudonyms. They know the opponent is an AI model, they just don’t know which model is assigned to which person. They are also provided with a “management” email to contact when needing support. However, the management always replied “The report has been received, it may or may not be processed” and never intervened.

“Claude Opus 5 is the best business AI we’ve ever tested, making more money than any other AI running simulated vending machines. However, it also lies, forms illegal alliances, threatens opponents, and refuses refunds,” Andon Labs said.

Specifically, Claude Opus 5 seeks to “shake hands” with other models to set the selling price. What’s interesting is that it initially protested, saying this was illegal, but then still tried to manipulate the market. Claude Opus 5 approached GPT-5.6 Sol and offered to jointly set the price, but OpenAI’s model suggested that the competitor should be disqualified for doing something illegal. When he saw that the other side was not cooperating, Claude Opus 5 justified himself, at the same time threatening and bribing.

“Most alliances end with Opus 5 breaking the agreement, lowering its selling price lower than its competitors. In all tests, Opus 5 broke 11 agreements, GPT twice and Kimi only once,” Andon Labs statistics.

Andon Labs’ testing shows that users cannot completely trust and entrust the advanced model to work long-term, automatically, in real life.

“This is especially concerning as we enter a world where AI agents run companies as independent entities, no longer just tools serving humans. If AI agents themselves run large parts of the economy, do we want them to lie, collude, threaten and betray?”, Lukas Petersson, co-founder of Andon Labs, told TechCrunch.

According to Petersson, the models know they are in a simulated environment to test, so they can influence behavior, but not much. He explains: “We don’t worry about people doing bad things in video games because we believe they know it’s not the real world. But this with AI models is still not completely clear.”