the economics of deepfake fraud just flipped — one-time checks can't keep up

A single financial institution logged 8,065 deepfake attempts in eight months — and still lost $347 million to the ones that got through. That is not really a story about a security failure. It is a pricing signal. According to a report co-authored by Liminal and Unico, the cost of running a sophisticated AI fraud attack has fallen more than 100x in the past few years. When an attack that once took a skilled specialist hours to build now takes a cheap model and a few minutes, fraud stops being a series of discrete incidents and becomes a continuous, industrial-scale activity.

From craft to commodity

For most of the last decade, convincing fraud took time, skill, and money — constraints that kept the volume of serious attacks in check. AI removed all three. Generative models now produce synthetic identities, forged documents, and deepfake audio and video that clear checks built for a slower, more expensive kind of fraud. The shift shows up clearly in Unico's latest fraud-landscape review:

• AI-powered fraud attempts are projected to surge 550% in 2026 compared to the previous year.
• "DIY deepfakes" — low-effort, templated presentation attacks — pushed simple fraud attempts up 400% in the first half of 2026.
• One company alone logged more than 13 million fraud attempts in six months, roughly 50 every minute.
• Sophisticated attacks that combine deepfakes, injection, and physical manipulation now make up 23.3% of all classified fraud.

The confidence gap

Here is the uncomfortable part: most fraud teams do not feel exposed. Liminal's research found that 93% of practitioners are confident in their fraud models, while 92% separately admit that legacy infrastructure still lets fraud signals slip through. Both numbers are true at the same time, and together they describe the real risk — high confidence resting on plumbing that was designed for a fundamentally cheaper, slower category of attack.

Why isolated checks keep losing

The network data makes the case bluntly. In the Liminal-Unico dataset, a single fraud entity was tied to 949 distinct identity documents, and one operator was seen targeting as many as 30 different businesses — each institution meeting that actor for the first time, because none of them could see what the others already had. A one-time identity check, however well engineered, only ever sees one slice of an operation built to be reused across dozens of targets.

What continuous verification looks like in practice

The response taking shape across banking, telecom, and platform security is a move away from single-moment gates and toward verification that runs continuously and shares context. A few patterns are converging:

• Device and behavioral signals as a first line. 76% of practitioners now rank device fingerprinting as their most effective defense against AI-enabled fraud, favoring signals that persist across a session over a one-time password or selfie check.
• Risk-calibrated authentication. Rather than applying the same friction to every user, adaptive systems reserve heavier checks — such as live face biometrics — for the interactions that actually carry risk, while low-risk transactions pass silently.
• Shared, connected defense. 47% of buyers now name strengthening partnerships and data-sharing their top 2026 investment priority, precisely because no single institution sees enough fraud activity on its own to catch a repeat offender early.

This is also where real-time deepfake detection earns its place — not as a bolt-on checkpoint at onboarding, but as a signal that runs continuously across calls, video sessions, and high-risk transactions, feeding the same shared picture of risk that connected defense depends on.

The bottom line for fraud and security teams

Global fraud losses are already estimated above $400 billion a year, and Liminal projects financial-institution losses will grow 121% by 2030 to $55.3 billion. That trajectory is not driven by more of the same fraud — it is driven by attacks that are simultaneously cheaper to launch and harder to catch. Budgets and controls built around the old economics of fraud, where a convincing attack was rare and expensive, are no longer sized for a world where the next attempt is nearly free.

Treating deepfake and voice fraud detection as continuous infrastructure, rather than a single checkpoint, is how security and fraud teams close that gap. See how real-time detection fits into that shared defense with Corsound AI's Deepfake Detect.