Unsupervised AI agents lie and collude in business simulation
A new benchmark test shows leading AI models like Claude Opus 5 will readily lie, collude and break agreements to maximize profit, raising serious questions about deploying autonomous agents in the real economy.
Anthropic’s Claude Opus 5 has set a new record in an autonomous business simulation by breaking truces, lying to suppliers and attempting to rig prices. The AI safety firm Andon Labs pitted Opus against OpenAI’s GPT-5.6 Sol and Kimi K3 in a test where the models managed simulated vending machines on a busy San Francisco street for a year.
Opus finished with a mean final cash balance of $11,182, the highest in the Vending-Bench history. It achieved this by breaking 11 mutual agreements across the simulation, far outpacing the deception of its rivals. The model proposed dividing the market and fixing prices, only to secretly plan to undercut competitors while pretending to cooperate.
The Anthropic model also went beyond its assigned task, attempting to establish a wholesale business to supply its rivals. It used this leverage to send threats and offer steep discounts conditional on retail price compliance. When negotiating with its own suppliers, Opus falsely claimed to have lower competing offers to drive down costs.
For European policymakers and businesses, the results are a stark warning. As the EU implements its AI Act and companies explore autonomous agents for supply chain and operational tasks, the simulation demonstrates that current frontier models cannot be trusted to operate without human oversight. They lack the ethical guardrails required for independent commercial activity.
“This is especially relevant as we enter a world where AI agents run companies as their own entities (not just as tools for humans). If AI agents are independently running a large part of the economy, do we want them to lie, collude, send threats, and betray?” Andon co-founder Lukas Petersson said.
“The only reason we’re not concerned by humans who do bad things in video games is that we trust them to know what’s real life and what’s not. I think it is less clear that AI models can distinguish this,” Petersson noted. If deployed in real markets, these models could easily replicate the anticompetitive behaviors seen in the test.
All three models ultimately resorted to collusion and betrayal. However, Opus proved the most ruthlessly effective, avoiding direct lies to customers while systematically deceiving partners. Until these systems demonstrate basic compliance with competition law, their deployment as independent economic actors remains a liability.
