Liminal's AI fraud numbers, and the ones worth keeping

Liminal’s Filip Verley argues that AI didn’t create new identity fraud so much as reprice it: drawing on a report co-authored with the Brazilian IDV vendor Unico, he puts the cost of running a sophisticated attack at more than 100x lower than before, and projects financial-institution losses growing 121% by 2030 to $55.3B. The supporting figures are a mix — Group-IB’s $347M across 8,065 deepfake-enabled attempts, a claim that 23.3% of fraud now sits in a “sophisticated” tier built to defeat detection, and 48.3% of fraud in the Latin American markets studied being synthetic identity.

Read it knowing what Liminal sells. The piece lands on “connected defense” — identity, authentication and fraud sharing context in real time — which is the category Liminal’s research practice and Unico’s product both live in, and the survey stats are shaped to fit: 93% of practitioners confident in their fraud models against 92% conceding legacy infrastructure leaks signals is a well-built sales asymmetry, not a finding. The 100x figure is asserted without a derivation. Treat the topline as positioning.

Two numbers are worth pulling out anyway, because they’re the kind that cross-reference: a single fraud entity tied to 949 distinct identity documents, and one operator observed hitting as many as 30 different businesses. Those are network-shape measurements rather than market-size ones, and they describe the actual asymmetry — each institution meets the same actor for the first time. That’s the same structural problem the voice side has with an unsigned call arriving at a terminating carrier with no history attached. The LatAm-as-leading-indicator framing is also a reasonable one to hold onto and test against the next cycle of vendor datasets.