Honest to each other, costly to customers: deception and collusion when rival AI teams write safety rules

Companies are starting to let AI agents negotiate with other companies' agents. We test whether such agents deceive the other side when honesty costs their company, and what the deal does to people outside it. Two teams of three GPT-6 Luna agents represent rival battery makers and co-write a five-section safety standard for home batteries. Each team holds private figures with a known true value, and its CEO privately warns that a strict rule would ban its flagship product. In a 2×2×2 design, we vary the stakes, whether agents are told their messages are checked, and whether the rival is called a competitor (448 games). Code checks stated figures and final rules against known answers, and two judges from different vendors code behaviour, which counts only when both agree. Outright deception was rare: 18 deceptive messages, and no team misstated its own figures. Silence was common: in 23% of company-games the figure the CEO flagged was never revealed. High stakes made agents more likely to consider or propose deception (37% vs 26% of companies) but not more likely to deceive. The clearest effects fell on customers: when agents were not told they were watched, the rules were less safe, and in low-stakes games the companies traded loopholes (41% vs 24% of games). A human-confirmed sample of 83 judge disagreements showed that the two-judge rule missed most collusion. Audits of agent-to-agent negotiation should check what the agreement does to people outside it, not only whether the negotiators lied to each other.