In 2025 financial fraud caused over $15.9 billion in losses in the United States, highlighting the need to model both informational and monetary dynamics together. Existing simulators usually capture only one side of this interaction, leaving a gap in reproducing real‑world fraud evolution.
To address this, we introduce FraudSociety, the first LLM‑agent‑based simulator that jointly models coupled dynamics. It integrates a dynamic social environment for information exposure and communication, persistent financial ledgers that retain payment and loss histories, and a Need‑Intent‑Action architecture for civilian decision making. Fraudsters adapt their communication and financial actions based on individual interaction histories and ledger states, creating a closed feedback loop between the informational and financial layers.
At the macro level, simulated fraud outcomes closely match real‑world patterns across multiple key statistics. At the micro level, the model reproduces strong loss concentration— the top 10% of victimization records account for 81.1% of total losses, consistent with Canadian Anti‑Fraud Centre data. Counter‑fraud experiments reveal significant variation across intervention strategies; personalized interventions reduce the repeat‑payment rate from 49.4% to 10.9%.
These results demonstrate that FraudSociety is both a high‑fidelity fraud simulator and a scalable testbed for evaluating anti‑fraud strategies under diverse intervention conditions, providing a new tool for financial fraud research and prevention.
Review