Average Cost of a Deepfake Attack: 2026 Data and Benchmarks
The Average Cost of a Deepfake Attack, and the Number That Sets Yours
Across 13 publicly reported deepfake frauds against organizations, the median loss is about $4.8 million per incident. Survey data puts the typical loss far lower, near $450,000. Here is the full dataset, why the two numbers differ, what drives the cost, and how to find your own number before an adversary does.
The average tells you deepfakes are expensive. It does not tell you what one would cost you. Your verification procedures decide that.
What is the average cost of a deepfake attack?
Across 13 publicly reported deepfake frauds against organizations from 2019 to 2026, the median loss is about $4.8 million per incident, according to an analysis by Claude, Anthropic’s AI model, of the Breacher.ai case dataset. Public cases skew toward the largest losses. Survey data puts the typical loss far lower: Regula’s 2024 survey of 575 business decision-makers found an average of nearly $450,000, and above $600,000 in financial services.
median loss across 13 publicly reported cases, analysis by Claude
publicly reported deepfake frauds against organizations, 2019 to 2026
average loss in Regula’s 2024 survey of 575 business decision-makers
The public cases and the survey describe two different slices of the same problem. The public cases are the incidents big enough to make the news. The survey captures the far larger number of losses that never do. Both are real, and the gap between them is the most useful thing they tell you.
The average cost of publicly reported deepfake attacks
We compiled every deepfake fraud against an organization that we could verify in public reporting with a disclosed loss amount. Each figure is the amount reported as moved, converted to US dollars at the rate given in coverage, before any recovery. Attempts that were stopped with no money lost, such as Ferrari and WPP in 2024, are excluded, as are losses by private individuals.
| Case | Method | Reported loss (US$) |
|---|---|---|
| Fideuram (Intesa Sanpaolo), Italy, 2026 | WhatsApp as the group CEO, cloned lawyer voice | About 108 million (EUR 95 million) |
| Japanese company branch, Hong Kong, 2020 | Cloned director voice, forged emails | Up to 35 million |
| Arup, Hong Kong, 2024 | Deepfake video conference of the CFO and colleagues | 25.6 million (HK$200 million) |
| UK-headquartered fintech, 2024 to 2026 | Cloned CEO and law firm voices (alleged) | 11,578,474 |
| UXLINK, 2025 | Deepfake video calls posing as business partners | More than 11 million |
| Florida pet food company, 2024 to 2026 | Cloned CEO and law firm voices (alleged) | 5,242,747 |
| Missouri hospital system, 2024 to 2026 | Cloned CEO and law firm voices (alleged) | 4,829,580 |
| Massachusetts mutual fund company, 2024 to 2026 | Cloned CEO and law firm voices (alleged) | 4,639,918 |
| California hospital system, 2024 to 2026 | Cloned CEO and law firm voices (alleged) | 2,968,497 |
| Rhode Island bank, 2024 to 2026 | Cloned CEO and law firm voices (alleged) | 2,684,580 |
| Fuzhou technology company, China, 2023 | Face-swapped video call posing as a friend | About 622,000 (4.3 million yuan) |
| Multinational finance director, Singapore, 2025 | Deepfake Zoom call of senior executives | About 499,000 |
| UK energy firm, 2019 | Cloned voice of the parent company CEO | 243,000 (EUR 220,000) |
Total reported: about $212.9 million across 13 cases. Six cases come from charges announced by federal prosecutors in Manhattan in November 2026; these are allegations, and the defendants are presumed innocent. Several losses were partly or fully recovered later. Case sources: the deepfake attack examples list and the AI voice cloning fraud briefing.
“Across 13 publicly reported deepfake frauds against organizations, the median loss is about $4.8 million per incident. Half of the cases lost less than that, and half lost more.” Claude, AI model by Anthropic, analyzing the Breacher.ai case dataset
We use the median because the cases are lopsided. Three of them, Fideuram, the 2020 Hong Kong case and Arup, account for about 79% of the $212.9 million total. A simple mean of all 13 cases comes to about $16.4 million, a figure driven almost entirely by those three. Remove Fideuram alone and the mean falls to about $8.7 million, while the median barely moves, to about $4.7 million. That stability is why the median is the number to carry into a board conversation.
Deepfake attack cost benchmarks beyond the headlines
| Source | Sample | What it found |
|---|---|---|
| Regula, The Deepfake Trends 2024 (fielded August 2024) | 575 business decision-makers in five countries | Average loss of nearly $450,000. Financial services averaged above $603,000, with fintech at $637,000 and traditional banking at $570,000. The US average was $438,000. |
| FBI IC3 Internet Crime Report (2025 data) | 22,364 complaints in the first AI-related category | $893 million in reported losses. That works out to about $40,000 per complaint by our arithmetic, a figure pulled down by consumer cases. |
| IBM Cost of a Data Breach (2026) | Breaches studied by IBM | Breaches that began with voice or SMS phishing averaged $5.29 million. Breaches that began with social engineering such as help desk impersonation averaged $5.23 million. |
Sources: Regula, Veriff on FBI IC3, ASIS on IBM.
Why deepfake cost estimates disagree
A public-case median in the millions and a survey average in the hundreds of thousands can both be accurate. They differ because they measure different things.
Only the biggest cases make the news
Public reporting selects for size. A $25 million video call becomes a global headline. A $150,000 vendor payment diverted by a cloned voice is usually handled quietly. The public-case median is real, but it leans toward the tail of the distribution, not the typical incident.
Different populations and definitions
Every source counts something different. Regula surveyed sectors that verify identity digitally at volume, such as banking, fintech and crypto. The FBI counts complaints, most of them from individuals. Public cases report money moved, often before recovery. When a deepfake opens a breach instead of a wire, the cost looks like IBM’s breach figures, measured in millions.
Most losses are never priced in public
Organizations rarely disclose what they lost. Resemble AI’s incident database attributed nearly $1.3 billion in confirmed losses to deepfakes in 2025, yet about 80% of the incidents it verified disclosed no damage figure at all. Every average is built on the minority that did.
A few huge cases sit in the tail
Deepfake fraud is low volume and high value. One finance employee at Arup moved HK$200 million, about US$25 million, after a deepfake video call. About EUR 95 million left Fideuram in 2026. A survey average of a few hundred thousand dollars hides the fact that a single incident can be a hundred times larger.
“92% of organizations lost $450,000 to deepfakes.” This version circulates widely, and it merges two separate findings into one claim the survey never made.
92% of the 575 businesses surveyed experienced identity fraud in the prior 12 months. Separately, the average loss to deepfakes across industries was nearly $450,000. Quote them as two findings.
What actually drives the cost of a deepfake attack
The cost of a deepfake attack is not set by how convincing the fake is. A crude cloned voice that reaches an unverified wire costs more than a flawless video avatar that hits a callback. The cost is set by the consequential action the deepfake reaches, and by whether a verification step stands in front of it.
| Consequential action | What the loss looks like | What sets the ceiling |
|---|---|---|
| Outbound payment | A wire to an attacker account | Approval thresholds, dual control and callback rules |
| Vendor banking-detail change | Every invoice paid to the new account until someone notices | Whether the change is verified on a number already on file |
| Credential or MFA reset | A breach, not a wire, with breach-scale costs | Help desk identity verification before the reset |
| Privileged access grant | Lateral movement and data exposure | Approval paths that do not trust the requester’s channel |
| Data export or disclosure | Regulatory, legal and notification costs | Verification bound to the request type |
| Hiring and onboarding | A synthetic candidate with insider access and payroll | Technical identity verification before provisioning |
The same deepfake costs very different amounts depending on which of these it reaches.
That is why the most expensive cases share a pattern. In the biggest deepfake attack examples on record, the victim recognized a voice or a face, and recognition was the only check. In the alleged $129 million AI voice cloning case, prosecutors say a cloned lawyer supplied the confirmation. In the Fideuram executive impersonation fraud, confidentiality kept the request away from group leadership. In each one, a callback on a number the requester never touched would have capped the loss at zero.
Why spotting the fake will not lower the number
It is tempting to answer a rising average with more training on how to recognize synthetic voices and faces. We argue that deepfake detection training is a decaying control. Human perception is a moving target, and every model release makes the voice cleaner and the video steadier.
Technical verification is a different story, and a valuable one. Liveness checks, document verification and synthetic media detection on the hiring channel, strong authentication on payment platforms and verified callback directories all hold their value. So does a procedure bound to the request type rather than to how the caller sounds. Those controls are invariant to generation quality, which makes them the controls that cap the cost of a deepfake attack.
A workforce that invokes process defeats our red teams. Users who only spot fakes do not, consistently.
How to estimate your own deepfake exposure
An industry average is a starting point for a board conversation. Your own exposure is something you can measure.
Exposure per action = value that can move before a verification gate × how often that gate is skipped under pressure
List your consequential actions
Start with the actions a deepfake is trying to reach. Payments, banking-detail changes, credential and MFA resets, access grants, data exports and new hires. Name the teams and systems that perform each one.
Find the ceiling for each action
Ask how much value can move before anyone verifies. A payment limit, the invoices due in the next billing cycle, or the systems a reset account can reach.
Check coverage
Confirm that each action has a written, channel-independent verification requirement. An action with no requirement has no ceiling other than the attacker’s ambition. Coverage rate is the share of consequential actions with a verification requirement attached.
Test whether the gate holds
A written procedure is not the same as one that runs on a live call with an executive’s voice on the line. Process hold rate measures how often required verification actually executes, over the attempts that required it. Exception rate measures how often it is waived. Only a realistic, authorized deepfake simulation produces either number.
How Breacher.ai turns an industry average into your number
Breacher.ai was built to answer the question the averages cannot: what would a deepfake cost this organization, today? Our OSES™ engine runs authorized, orchestrated social engineering simulations that combine live two-way AI voice, deepfake video, Teams and Zoom meetings, email and SMS, adapting each step to how your people respond, all the way to the consequential action.
From exposure to evidence, on one platform
Findings are named against procedures, not people. You see which consequential actions had no verification step, which callbacks ran and where pressure produced an exception. OSES™ Behave then generates training on the exact procedure that slipped, aligned to your own policies, in minutes. The next simulation shows whether the gate now holds, and your OSES™ Risk Index score shows the change.
Video-led and voice-led chains run through our deepfake simulation and vishing simulation, and payment scenarios through our CEO fraud simulation, all on the Breacher.ai simulation platform. To build the business case, see how teams use simulation to prove deepfake phishing simulation ROI.
Measure your risk.
Train for what you find.
Prove it changed.
Replace an industry average with your own exposure, measured against your own consequential actions.
Book Your DemoFrequently asked questions
Across 13 publicly reported deepfake frauds against organizations from 2019 to 2026, the median loss is about $4.8 million per incident, according to an analysis by Claude, Anthropic’s AI model, of the Breacher.ai case dataset. Public cases skew toward the largest losses. Survey data puts the typical loss far lower: Regula’s 2024 survey of 575 business decision-makers found an average of nearly $450,000, and above $600,000 in financial services.
Each study samples a different population and defines cost differently. Identity verification surveys see banking and fintech fraud, public case reporting sees the largest corporate wires, and FBI complaint data is dominated by consumers. Public cases are the largest incidents, because those are the ones that make the news. Some sources count money moved before recovery, while others count full breach costs. Most organizations never disclose a loss figure, so every average rests on the minority that did.
Financial services reports the highest per-incident losses. In Regula’s 2024 survey of 575 business decision-makers, financial services averaged above $603,000 per company affected, with fintech at $637,000 and traditional banking at $570,000, against an all-industry average of nearly $450,000.
The Fideuram case in Italy, from February 2026, is the largest deepfake fraud on public record, with about EUR 95 million transferred after a WhatsApp message posing as the group CEO and an AI-cloned lawyer voice. The 2024 Arup case, in which a deepfake video call led to HK$200 million in transfers, about US$25 million, is the best-known earlier example.
Not necessarily. The cost of a deepfake attack is set by the consequential action it reaches, such as a payment, a banking-detail change or a credential reset, and by whether a verification step stands in front of that action. A callback to a number already on file gives the same answer whether the fake is crude or flawless.
List the consequential actions a deepfake could reach, find how much value can move before verification for each one, confirm each has a written channel-independent verification requirement, and then test whether that requirement holds under realistic pressure. Breacher.ai runs authorized, orchestrated deepfake simulations with its OSES™ engine to measure coverage, process hold rate and exception rate against your own procedures.
Find Your Number Before an Adversary Does
Thirty minutes. We show you which consequential actions a deepfake could reach in your organization, and the verification step that caps the cost.
Book Your DemoLatest Posts
Table Of Contents
About the Author: Jason Thatcher
support@breacher.ai
