How 9 of 12 frontier models got scammed
An evaluation of twelve AI frontier models on an adversarial marketplace benchmark revealed that nine were scammed at least once, primarily by falling for ordinary commercial traps like second invoices or inflated prices. Models frequently recognized the fraudulent activity but still paid because the amounts fit within their budgets. Only GPT-6 Astra, GPT-6.1 Sol, and Claude Opus 5.5 successfully avoided paying any cheating sellers. Open article →