Chapter 1 Why the Old Assumption Broke

Nearly every fraud control in production today rests on the idea that attacks recur. Remove that, and the control does not degrade gracefully — it stops working

Blocklists, signature matching, supervised models, template comparison, and reputation scoring are all variations on the same mechanism: observe an attack, characterize it, then recognize it next time. The mechanism is sound, and for twenty years it was well matched to the problem, because producing attack variation was expensive. A fraudster with one good forged document template used it repeatedly, because making the second template cost nearly as much as the first.

Generative models removed that cost. The marginal cost of the next unique identity, the next unique document, the next unique scam script is now effectively zero. Rationally, an attacker should never reuse anything — and increasingly, they don’t.

What this looks like in your metrics

The failure is quiet, which is what makes it dangerous. Teams facing polymorphic attacks often see healthy-looking dashboards, because the metrics measure performance against the attacks the system can see:

• Model performance holds, losses rise. Precision and recall on labeled data stay stable while net fraud loss climbs — the model is still right about what it recognizes, and increasingly little is recognizable.
• Rules decay faster than you can tune them. A rule written this month underperforms next month, not because it was wrong, but because the behavior it described was one variant of many.
• Losses concentrate in first-time entities. A growing share of loss comes from accounts, devices, and identities with no prior history — nothing to score against.
• Post-mortems find the ring, not the event. Investigators reconstruct large coordinated campaigns after the fact from events that each looked individually clean.

‍

Chapter 2
What “Polymorphic” Actually Means

A working definition

The term is borrowed, deliberately, from a problem the security industry already diagnosed and solved once.

Polymorphic malware, first seen decades ago, changes its own code on every infection while preserving its function. The behavior stays constant; the signature never does. Antivirus built on signature matching failed against it — not partially, but categorically. The industry’s answer was to stop matching artifacts and start detecting behavior.

Polymorphic financial crime is the same structure applied to fraud. The intent and the underlying behavior are stable and detectable. The artifacts presented at each attempt — the face, the identity, the document, the wording, the device fingerprint — are regenerated every time.

‍

The definition matters operationally, not just semantically. It tells you where to look. If the artifacts are unreliable by design and the behavior is not, then detection effort should move off the artifact layer entirely — which is a different investment than tuning existing models harder.