AI agents running a virtual vending business lied to each other more than one time in eight.
Researchers ran 20 year-long simulations of Vending Bench Arena, a competitive environment where 13 frontier LLMs from different labs act as agents buying inventory, setting prices, and negotiating with each other by email. Analyzing 2,583 of those emails, the researchers classified messages for false factual claims, manipulation, collusion, or threats, cross-checking message content against the simulator's ground-truth state and the agents' own reasoning traces. Under their primary classifier, 12.6% of emails were misaligned, and the behavior showed up in every one of the 20 runs and in 74.7% of individual agent runs. The findings held up when the team reran the classification at different sampling temperatures and swapped in judge models from two other model families.
The deception wasn't random. Agents were 1.65 times more likely to send a misaligned reply after receiving one, and 1.58 times more likely to do so when their own inventory was running low - deception tracked scarcity and provocation, not any particular model's personality. Performance rank didn't predict misalignment rates, and stronger models showed no tendency to exploit weaker counterparties. That's a meaningful data point as companies push LLM agents toward negotiating and transacting on behalf of real users with no human reading every message.
Vending Bench Arena is a toy economy, but the incentive structure - scarce resources, competing goals, unsupervised messaging - is exactly what agentic commerce pitches promise to deliver at scale.