Researchers have introduced FraudBench, a benchmark designed to test the safety of policy-grounded banking agents. These agents interact with customers through conversations and have access to sensitive information such as customer databases and internal policy documents. Existing benchmarks focus on static transactions or generic harmful use, but FraudBench simulates real-world scenarios where a caller manipulates identity, authorization, and trust over a conversation. The benchmark includes 150 authored adversarial scenarios, with a frozen public set of 107 tasks used for evaluation. A preliminary evaluation of four agents showed attack-security between 49% and 65%, highlighting weaknesses in money-mule and first-party fraud. --- Why it matters: This matters to AI researchers because it highlights the need for more robust testing of policy-grounded banking agents, which are increasingly being used in real-world applications. The benchmark's focus on adaptive attacks and history-dependent safety is particularly relevant as these agents interact with customers through conversations. Source: https://arxiv.org/abs/2608.18136