Key takeaway: Commoditized dark-web fraud toolkits now let criminals defeat static, rule-based fraud checks at scale. AWS argues banks should add models trained on long behavior sequences and account networks to keep up.
AWS Head of Payment Networks Paul Chang told PYMNTS that banks are moving beyond static, rule-based fraud detection. He pointed to what AWS calls transaction foundation models, or TFMs. These models score long sequences of financial activity instead of judging the attributes of a single payment. Chang said TFMs are meant to complement banks’ existing machine-learning systems, not replace them.
Chang said fraudsters now buy commoditized toolkits on dark web marketplaces. These toolkits let them run scams, card testing, and synthetic identity attacks at industry scale. Static rule sets struggle to keep up with that volume. Chang described transaction foundation models as “large-scale AI models pre-trained on billions of financial events.” These models spot patterns across a customer’s activity over time, rather than flagging individual transactions in isolation.
Chang said fraud rings spread their activity across multiple accounts, identities, devices, and IP addresses to hide what they’re doing. Graph neural networks can treat accounts, transactions, and devices as connected nodes and map how they relate to each other. IP addresses and shared devices show up as relationships between those nodes — connections that can surface a fraud ring that single-transaction scoring would miss.
AI agents are also starting to help with investigations. They pull evidence from device records and customer profiles, then draft case summaries that cite their sources. Chang said human analysts still keep final authority over case dispositions and suspicious activity report filings.
The article cited two production results. Stripe’s Radar product cuts fraud by 32% on average. Razorpay reported identifying five times more fraudulent or disputed transactions without raising its overall alert volume.
Why it matters: Behavior-sequence models and fraud-ring graphs could help banks catch more organized fraud without raising false-alert rates. If that holds up, merchants and processors working with those banks could see banks decline fewer legitimate transactions while still catching more coordinated attacks, such as synthetic identity fraud.
Source: PYMNTS