Trust and Fraud Prevention in Autonomous AI Payments: Integrating Behavioral Anomaly Detection with a Mandate-Based Authorization Framework

Autonomous AI agents are beginning to initiate payments for people and firms, yet fraud controls and Know-Your-Customer rules still assume that a human initiates each transaction. We propose a Know-Your-Agent (KYA) architecture that combines mandate-based authorization, modeled on the Agent Payments Protocol (AP2) and x402, with behavioral fraud scoring in which an autoencoder anomaly score is fused into an XGBoost classifier. Because no public data on agent-initiated payments exist, we built a seeded simulator of 250,000 payments from 4,000 agent-principal pairs, with a 3.0% fraud rate distributed across five agent-specific attacks, and evaluated every model on pairs held out from training. Behavioral scoring raised F1 from 0.303 for mandate checks alone to 0.774, although the autoencoder added nothing measurable over XGBoost alone (F1 0.772). Removing the five mandate-context features lowered F1 to 0.63. In leave-one-attack-type-out tests, supervised recall on an unseen attack fell to a range between 0.3% and 27%, whereas the autoencoder used as a separate channel still detected 95% of unseen token replays and 46% of unseen prompt-injection payments. On the synthetic PaySim and Bank Account Fraud (BAF) benchmarks, a simulated mandate overlay doubled F1 on BAF only because it is generated from the fraud label, and zero-shot transfer between the two datasets was weak with the overlay and weaker without it. We conclude that mandate rules, supervised scoring and anomaly detection should run as parallel channels, and we set out the liability rules, monitoring thresholds and consent-record standards that a KYA framework requires.