Voice Is No Longer Proof. Behavior Still Is.

Something shifted under our feet this year. INTERPOL's 2026 threat assessment named impersonation fraud one of the leading contributors to more than $400 billion in global losses, and deepfake-driven phishing has climbed over 300% in the past two years. For those of us who build the defenses, the uncomfortable engineering takeaway is blunt: every signal we used to treat as proof - a voice, a face, a caller ID, a one-time code read aloud over the phone - can now be manufactured on demand, cheaply, at scale.

The Authentication Era Is Ending

For twenty years, fraud defense mostly meant proving identity harder: passwords, then device fingerprints, then biometrics, then step-up codes. Every layer quietly assumed the attacker was an outsider trying to get in. Authorized-push-payment scams and deepfakes break that assumption completely. The legitimate customer is logged in, verified, on their own trusted device - and committing the fraud themselves, under someone else's direction.

You cannot authenticate your way out of a problem where authentication already passed. When the real account holder is the one pressing send, every gate you built to keep strangers out waves them straight through.

What Can't Be Faked at Scale

So we went looking for the signal that survives when identity stops being trustworthy. The answer is behavior - not who you are, but how you act while you do something. The cadence of your typing, the pause before you confirm a payment, the way you move between fields, whether you're following your own intent or a voice in your ear telling you what to enter next.

A cloned voice is a recording of an output. Behavior is a live process - generated fresh in every session and shaped by the pressure you happen to be under at that moment. That makes it expensive to fake convincingly, and effectively impossible to fake at the volume industrialized fraud operations need to be profitable.

The Hard Part Is the Millisecond Before the Send Button

Reading behavior only matters if you can act on it inside the window that counts - between the moment a customer commits to a transfer and the moment the money actually leaves. That is an engineering constraint before it is a data-science one. It means scoring a full behavioral session in real time, under peak production load, at a 99th-percentile latency the payment flow can absorb without anyone noticing - and returning a decision the bank can still defend to an examiner months later.

We build backwards from that deadline. The model, the feature pipeline, and the intervention itself all have to fit inside the same tight budget, because a verdict that arrives after the funds have settled isn't prevention. It's a report.

Baselines, Not Blocklists

The other half of the problem is knowing what normal looks like for one specific person. A rule fires the same way for everyone; a baseline is personal. We build a behavioral fingerprint from a customer's own legitimate sessions, so the question is never "does this look risky in general" but "does this look like them, right now."

That distinction is what lets the system hold a coached transfer while letting a genuine one through untouched. The deviation is measured against the individual, not a population average - which is exactly why it can catch a manipulated session that passes every conventional check.

The Next Front: Agents Acting on Our Behalf

The same generative wave that made deepfakes trivial is now placing AI agents inside the payment flow - initiating and approving actions for customers, and increasingly targeted as the softest way in. The industry is converging on a "know your agent" posture: every automated action traceable to a verified human authorizer, auditable and identity-bound.

Behavioral intelligence extends here more naturally than most defenses do. An agent has a behavioral signature too - a rhythm and shape to how it operates - and a hijacked, spoofed, or manipulated one drifts away from it just like a coerced human does. The layer that learned to read people will have to read the agents acting for them. That's the problem we're building for next.