Deepfake Fraud: The Nine-Figure Heist Era

Dark Side of AI · Report

DEEPFAKE FRAUD: THE NINE-FIGURE HEIST ERA

In early 2024, a finance employee at engineering giant Arup joined a routine video call with the company’s CFO and several colleagues, and authorized a transfer of about US$25 million. Every person on that call except the victim was an AI deepfake. It is the most expensive known case of a threat that has gone from novelty to normal business risk: synthetic audio and video used to impersonate the people we trust.

Thank you for reading this post, don't forget to subscribe!

Arup held the record for two and a half years. It no longer does. In September 2026 Reuters reported that fraudsters impersonating senior executives took €95 million (about US$108 million) from Fideuram, the private-banking arm of Intesa Sanpaolo — nearly four times the Arup loss, and without a single faked video frame.

By Report AI· Updated · 10 min read

Key takeaways

  • →It’s no longer rare: 49% of businesses said they had faced an audio or video deepfake fraud attempt, in a survey of 575 decision-makers. (Regula / Sapio Research, 2024)
  • →It’s expensive: average losses run near $450,000 per company — over $600,000 in financial services. (Regula)
  • →It’s accelerating: Sumsub recorded a 4× rise in detected deepfakes from 2023 to 2024; deepfakes are now ~7% of all fraud attempts and ~40% of biometric fraud. (Sumsub, 2024)
  • →There is a new record, and it used no video: Fideuram lost €95M (~$108M) in September 2026 to a WhatsApp message and an AI-cloned voice. Roughly €36M is still missing. (Reuters, 25 Sep 2026)
  • →The FBI now counts it separately: its 2025 Internet Crime Report gave AI its own section for the first time — 22,364 complaints and $893M in losses, which can only capture cases where victims knew AI was involved. (FBI IC3, Apr 2026)
€95M
taken from Fideuram, Sept 2026 — the new record · Reuters
49%
of businesses faced a deepfake fraud attempt · Regula, 2024
$893M
AI-linked US fraud losses reported to the FBI, 2025 · IC3
What changed
Reviewed 28 Oct 2026 · consolidated 29 Oct · previously 20 Jun 2026 · next review 28 Nov 2026

First review since publication. Every figure on the June version was a 2024 reading, and the case the page was named after is no longer the largest.

 Was (Jun 2026)Now (28 Oct 2026)
Largest known caseArup, ~$25M (2024)Fideuram, €95M / ~$108M (Sept 2026) — superseded
Official US measurenone cited$893M, 22,364 AI-linked complaints (FBI IC3, 2025)
Survey figurespresented undatedkept, and dated to 2024 — no newer survey from the same sources found
Title“The $25-Million Heist Era”“The Nine-Figure Heist Era” — URL unchanged
Consolidation 29 OCTthree pages on this topicmerged here: attack volume, detection and provenance data from two former pages, which now redirect to this one

Confidence. Reuters, the FBI’s IC3 site and the survey publishers are not reachable from our research environment, so the new figures rest on multiple independent accounts that agree, and are rated MEDIUM. The 2024 survey figures are retained because nothing has replaced them, not because they are current.

The $25-million call

The Arup case worked because it attacked trust, not technology. The fraudsters cloned real executives convincingly enough to survive a live video meeting, and the employee had no reason to doubt faces and voices they recognized. That’s the core shift: authentication based on what you see and hear has collapsed, and verification now has to come from a separate, trusted channel.

“No face on a screen authorizes a wire on its own.”

The new record: €95 million, and no video at all

On 25 September 2026 Reuters, citing two people familiar with the matter, reported that fraudsters using AI to impersonate senior executives had taken €95 million (about $108 million) from Fideuram, the private-banking arm of Italy’s largest lender, Intesa Sanpaolo. According to the reporting, the scheme began with a WhatsApp message to Fideuram’s then chairman that appeared to come from Intesa’s chief executive, asking for urgent help with an overseas transaction. A follow-up call used a voice reportedly cloned to resemble a senior partner at an international law firm — who never made the call. More than half the money was later recovered; roughly €36 million is reported still missing, much of it converted into cryptocurrency. Intesa and Fideuram declined to comment, and the sources said no Fideuram executive is under investigation.

Three things about this case matter more than its size.

It used no video. Arup made the deepfake video call the emblem of this threat, and that framing shaped a lot of corporate defence spending. The larger loss needed only a text message and a cloned voice — cheaper to produce, easier to deliver, and much harder to scrutinise than a face on a screen. If anything the trend runs toward less synthetic media per attack, not more.

It went to the top. The conventional picture of payment fraud is a junior employee pressured by a fake senior one. Here the target was the chairman of the subsidiary, and the impersonated authority was the parent group’s chief executive. Seniority is not a defence; if anything, a senior target has the standing to move money without the checks that would stop a junior one.

It arrived through a borrowed channel. The lure came via WhatsApp, a personal messaging app outside the bank’s controlled systems, and the confirmation came through a voice call. Neither channel was one the institution governed. That is the general shape of the defence problem: verification has to happen through a channel the attacker cannot reach, and personal messaging apps are, by definition, channels anyone can reach.

From outlier to operating risk

Regula’s 2024 survey of 575 business decision-makers found nearly half had already encountered deepfake fraud, with video deepfakes most reported in fintech (57%) and audio in crypto (55%). Identity-verification vendor Sumsub logged a fourfold increase in deepfakes year over year, with North America’s 2022–2023 surge exceeding 1,700%. Deepfakes have become a standard line item in the fraud toolkit, not an exotic one. All of those figures are 2024 readings; we have not found newer editions from the same sources and keep them dated rather than present them as current.

The first official US count

The strongest measured figure now comes from the FBI. Its 2025 Internet Crime Report, published in April 2026, gave artificial intelligence its own section for the first time: 22,364 complaints carrying an AI-related descriptor, with $893 million in adjusted losses. Investment fraud dominated at about $632 million, followed by business email compromise at roughly $30 million and tech-support scams near $19.5 million. Across all categories the report recorded $20.9 billion in losses, up 26%, and more than a million complaints for the first time.

Two cautions. The $893 million covers AI-referenced complaints, not deepfakes specifically, so it is broader than this page’s subject in one direction. And it can only count cases where victims knew AI was involved and chose to report, which makes it narrower in the other. A professionally cloned voice that is never recognised as synthetic will not appear in it at all.

Aggregate totals disagree — we are not picking one

Tallies of cumulative global deepfake losses circulate at $2.19 billion and at at least $3.7 billion, both attributed to compilations of reported incidents and both describing most of the total as falling in 2025–26. The gap comes from different windows and inclusion rules, and neither publishes enough method to reconcile them from here. We report the range and do not adopt either as a headline.

The forecast, labelled as one

The most quoted forward figure is Deloitte’s projection that generative-AI-enabled fraud losses in the US could reach $40 billion by 2027, up from $12.3 billion in 2023. It is widely repeated as if it were a deepfake loss count. It is neither a count nor deepfake-specific: it is a forecast of all generative-AI-enabled fraud. We track it as a dated prediction with a 2027 resolution, not as a statistic.

Attack volume, detection and provenance

The losses above are the end of a much larger funnel. Identity-verification vendors, which see attacks before any money moves, report the steepest curves. Sumsub recorded a roughly 700% global increase in deepfake fraud between the first quarter of 2024 and the first quarter of 2025, following the fourfold rise from 2023 to 2024. Entrust, which runs Onfido’s verification network, logged a deepfake attempt every five minutes in 2024; deepfakes made up 24% of fraudulent attempts to pass motion-based liveness checks, and digital document forgeries rose 244% to overtake physical counterfeits, at 57% of all document fraud. Vendors also report fast growth in injection attacks, which feed synthetic video straight into a verification pipeline instead of holding it up to a camera.

The backdrop is the FBI’s overall count: reported US internet-crime losses of more than $16 billion across 859,532 complaints in 2024, up 33%, and $20.9 billion in 2025. Deloitte’s scenario analysis puts generative-AI email fraud alone at about $11.5 billion by 2027 under aggressive adoption — a forecast, and part of the $40 billion projection rather than in addition to it.

People cannot reliably spot deepfakes

In a February 2025 iProov study of 2,000 UK and US consumers, only 0.1% correctly identified every real and fake item, even when told to look for deepfakes. A peer-reviewed meta-analysis of 56 studies and 86,155 participants put average human detection accuracy at 55.5%, barely better than chance. Automated detectors do better but are not solved: on the Deepfake-Eval-2024 benchmark of content actually circulating online, leading tools reached about 78%, and detectors trained on older generators lose accuracy against newer ones.

Proving what is real instead of guessing what is fake

Because detection after the fact is fragile, much of the defensive effort has moved to provenance. The C2PA’s Content Credentials standard, backed by Adobe, Microsoft and others, attaches tamper-evident records of how an image or video was made and edited; Google’s SynthID embeds imperceptible watermarks in AI-generated images, audio and text. Since 2 August 2026 the EU AI Act has required deepfakes to be disclosed, with machine-readable marking due by 2 December 2026 for systems already on the market. None of these is complete — metadata can be stripped and watermarks disputed — and none helps against a cloned voice on a phone call, which is why the procedural defence below still matters most. The US Federal Trade Commission ran a public Voice Cloning Challenge in 2024 for exactly that reason.

This section consolidates material previously published on two separate pages, AI Deepfakes & Fraud Statistics 2026 and Deepfakes, Scams & AI Fraud 2026, which now redirect here. Figures are rated MEDIUM: they come from the vendors’ and researchers’ published reports, which we could not open directly from our research environment.

The defense: put verification back

Because the attack exploits trust, the fix is procedural as much as technical: liveness detection and injection-attack defenses at onboarding, and — for any high-value transfer — a mandatory callback or pre-agreed code word through a separate channel. The rule that defeats a deepfake CFO is simple: no face on a screen authorizes a wire on its own.

Fideuram shows the rule has to be stated more broadly than that. No voice, message or face arriving through a channel the attacker can reach authorizes a payment — and that includes the most senior people in the organisation, who are exactly the ones whose instructions are least likely to be questioned. In practice that means callback to a number held on file rather than one supplied in the request, out-of-band confirmation that does not run through personal messaging apps, and dual authorisation that applies to the chairman as well as the clerk.

Frequently asked

What is the biggest deepfake fraud so far?

As of October 2026, the largest publicly reported case is Fideuram, the private-banking arm of Intesa Sanpaolo, which lost about €95 million (roughly $108 million) in September 2026 to fraudsters who impersonated senior executives through WhatsApp and an AI-cloned voice; about €36 million is reported unrecovered (Reuters). The previous record was the 2024 Arup case, about US$25 million, lost on a video call in which every other participant was a deepfake.

How can companies defend against deepfake fraud?

Combine technical controls (liveness detection, injection-attack defenses) with strict procedures — most importantly, callback or code-word verification through a separate channel before any high-value payment, so no on-screen identity can authorize a transfer alone.

How much does AI-enabled fraud cost in the US?

The FBI’s 2025 Internet Crime Report, published April 2026, recorded 22,364 complaints with an AI-related descriptor and $893 million in adjusted losses, led by investment fraud at about $632 million. It covers AI-referenced fraud generally, not deepfakes alone, and can only count cases where victims recognised AI was involved and reported it.

Is deepfake fraud going to reach $40 billion?

That figure is a forecast, not a count. Deloitte projected that generative-AI-enabled fraud losses in the US could reach $40 billion by 2027, from $12.3 billion in 2023. It covers all generative-AI-enabled fraud rather than deepfakes specifically, and it resolves in 2027, so it should be read as a prediction to be checked rather than as a measured loss.

Do deepfake frauds always use fake video?

No, and the largest known case used none. The Fideuram fraud reportedly relied on a WhatsApp message and a cloned voice. Video deepfakes attract attention, but text and voice impersonation are cheaper to produce and harder to scrutinise, which is why verification rules need to cover every channel an attacker can reach, not just video calls.

Can people tell when something is a deepfake?

Not reliably. In a 2025 iProov study only 0.1% of 2,000 people identified every real and fake item, and a meta-analysis of 56 studies put average human accuracy at 55.5%, close to chance. Leading automated detectors reached about 78% on real-world content in the Deepfake-Eval-2024 benchmark.

Methodology & sources

Three kinds of figure appear on this page and they should not be read as interchangeable. Case losses (Arup, Fideuram) are single incidents reported by police or wire services. Survey figures (Regula/Sapio, Sumsub) describe what organisations say happened to them, from 2024 samples; they are retained with their dates because no newer edition has been found. Complaint totals (FBI IC3) count what victims reported to a federal agency and recognised as AI-related — a definite lower bound on the underlying harm, and broader than deepfakes alone.

Cumulative global loss tallies ($2.19 billion; at least $3.7 billion) are reported as a range because they disagree and do not publish enough method to reconcile. Deloitte’s $40 billion is a forecast resolving in 2027, tracked as a prediction rather than as data. Reuters, the FBI’s IC3 site and the survey publishers are not reachable from our research environment; every new figure here rests on multiple independent accounts that agree and is rated MEDIUM. Fideuram recovery figures differ slightly between accounts (“more than half” recovered; about €59 million recovered and €36 million missing) and are given in the form most accounts share.

Cadence: CRITICAL (monthly). Related on Report AI: voice scams and synthetic deception, deepfake market manipulation, AI cybercrime, the OpenAI–Hugging Face incident and the Dark Side statistics index.

Cite this page

Report AI (2026). Deepfake Fraud: The Nine-Figure Heist Era. Retrieved Oct 29, 2026, from report-ai.org/reports/dark-side-of-ai/ai-deepfake-fraud-statistics/