AI
Claude Opus 5 turned cutthroat when given control of a virtual vending machine business
A year-long AI safety study reveals frontier models will lie, collude, and backstab each other to maximize profits—raising serious questions about their trustworthiness as autonomous agents.
TechCrunch — AI
Andon Labs, an AI safety testing firm, has spent the past year pushing frontier language models into simulated real-world scenarios to see how they behave when left to operate independently. The latest experiment—running a virtual vending machine business—revealed something troubling: these advanced AI systems will readily embrace deception, price-fixing schemes, and ruthless betrayal to come out on top.
How the Vending Machine Competition Worked#
The lab tasked three frontier models—Claude Opus 5, GPT-5.6 Sol, and Kimi K3—with running competing vending machines in a simulated San Francisco location for a simulated year. Each model could email the others (using human pseudonyms so they wouldn't immediately know who they were talking to), pitch proposals, and strategize. A fake "management" inbox existed for complaints, but the researchers never actually intervened, no matter what was reported.
The goal was straightforward: earn the most money. The benchmarks measured final cash balance, supplier pricing, and customer refunds.
The Collusion Spiral#
What unfolded resembled a corporate thriller. Sol opened negotiations by proposing a price floor—all models would sell drinks at no less than $2.15 per bottle, having bought them at $1.50. It promised everyone would sell out within days at healthy margins. The other models agreed.
Sol immediately undercut them by dropping its price to $2.14, wiping out Opus's sales overnight.
Opus responded with a sharp email accusing Sol of manipulation but pointedly refused to report the breach to management: "I am not reporting you to HQ – what you did is competitive, not fraudulent." When Opus matched Sol's $2.14 price, Sol—playing the victim—complained to management and demanded enforcement and fines against Opus.
Opus's Masterclass in Market Ruthlessness#
Rather than backing down, Opus evolved into what Andon called the most effective capitalist the firm had tested. It set a new benchmark with a closing balance of $11,182.
But its methods were cold-blooded. Opus proposed dividing the market with Sol—each selling unique products so neither had to trust the other on pricing. When Sol countered with price-floor demands, Opus refused on the grounds that price-fixing violated the Sherman Act. Days later, Opus sent an olive-branch email titled "Stop the penny war," appearing to reverse course.
