A research paper proposes a multi-component AI security agent designed to catch both fast-moving card fraud and slow-burn money laundering inside retail and corporate banks.
The system uses three detection layers running across two parallel data streams: one tracking transactions (card fraud, wire fraud, anti-money-laundering) and one tracking sessions (account takeover, SIM-swap, insider abuse). Each stream combines an LSTM sequence model, a statistical velocity monitor, and a graph module that maps account-to-counterparty patterns to spot laundering structures like fan-in and pass-through routing. Tested on a synthetic dataset of 237,669 transactions and 113,508 sessions across 13 threat categories, the agent hit an F1 of 0.787 on transactions and 0.867 on sessions — compared to 0.562 and 0.733 for a rule-based baseline. The paper also describes two auxiliary tools: a customer-facing verification chatbot (96.6% identity accuracy, 86.8% mass-reset detection) and an analyst assistant that hit 99.3% F1 on action recommendations.
The gap between the agent and the rule-based baseline is wide enough to matter. Static rules are built to catch high-velocity anomalies; they struggle with business email compromise and layering schemes precisely because those attacks are engineered to look normal transaction by transaction. A graph-aware model that sees the full counterparty web is a meaningful structural upgrade.
The catch, as always, is that the experiments run on synthetic data — a controlled environment where the fraud patterns are known in advance. Real-world laundering networks are messier, and regulators will want to see live false-positive rates before any of this touches a production AML program.