Deepfake fraud cases explained in one line: almost every confirmed case is an ordinary payment, loan or account-access request delivered through a cloned voice or face, and the loss lands when no out-of-band check exists. The best-documented case, Arup’s Hong Kong office in January 2024, lost HK$200 million (about US$25.6 million) after a fake video conference.
These are misuse incidents, not vulnerabilities: no CVE, no CVSS score, no patch. Attackers turn a generative capability against a business process, which places them in the misuse bucket of the incident taxonomy. The cases below rest on police statements, FBI and FinCEN advisories and named reporting, graded against source verification tiers.
What are the major deepfake fraud cases on record?
Eight cases with a named primary or independent source define the pattern: three Hong Kong police cases, the 2019 UK energy-firm voice clone, failed attempts against WPP and Ferrari, a Hong Kong deepfake romance syndicate, and an FBI-tracked campaign impersonating senior US officials. Two of them, against WPP and Ferrari, failed outright.
| Date | Target | Method | Outcome |
|---|---|---|---|
| Mar 2019 | UK energy firm | Cloned voice of parent-company CEO | €220,000 wired; third request refused |
| 2022–2023 | Hong Kong lenders | AI face-swap loan applications | HK$200,000 in loans; nine arrests |
| Jan 2024 | Arup, Hong Kong | Pre-recorded deepfake video conference | About HK$200 million lost |
| May 2024 | Multinational trade firm, Hong Kong | Altered video of the UK CFO | Told to wire nearly HK$4 million |
| Disclosed May 2024 | WPP | Voice clone in a Teams meeting | Attempt failed |
| Jul 2024 | Ferrari | Cloned CEO voice by phone | Stopped by a verification question |
| Oct 2024 | Romance-scam victims across Asia | Face-swapped video calls | About US$46 million; 27 arrests |
| 2023–2025 | Contacts of senior US officials | AI voice messages and texts | Account takeover; ongoing |
How did the Arup deepfake video call work?
Police believe the Arup call was pre-recorded, not live. According to the Hong Kong Secretary for Security’s June 26, 2024 written reply to the Legislative Council, a fraudster emailed a Hong Kong employee posing as the UK head office’s CFO and invited them to a video conference about confidential transactions, then played footage built from public clips and voice recordings.
There was no real interaction. After the employee authorized transfers to five local bank accounts, the fraudster ended the meeting on a pretext and kept issuing payment instructions by instant message. The loss was about HK$200 million, and the reply described the investigation as ongoing. The government did not name the company; Arup was later identified as the victim, and AI Incident Database Incident 634 aggregates 30 reports on the case.
The sequence matters for defenders. A phishing email carried the request, and the video conference served as the authentication step that made it look legitimate.
A second case followed the same script. On May 20, 2024, a staff member at a multinational trade company spent nearly 30 minutes in a video conference with a fake “boss” impersonating the UK CFO and was told to transfer nearly HK$4 million to a local account. Police believe the suspect played public online videos altered with deepfake technology.
Which deepfake fraud attempts failed, and why?
Three documented attempts failed, fully or in part. A Ferrari executive asked a question only the real CEO could answer, the 2019 UK victim refused a follow-up payment after a promised reimbursement never arrived, and an attempt on a WPP agency leader failed, according to reports collected by the AI Incident Database.
WPP, disclosed May 2024. Per Incident 983, which collects reports from The Guardian, the Financial Times and others, scammers created a WhatsApp account using chief executive Mark Read’s image, set up a Microsoft Teams meeting, and used a voice clone and YouTube footage to press an agency leader to set up a new business. The aim was money and personal data.
Ferrari, July 2024. As The Drive reported, citing Bloomberg, an executive received WhatsApp messages about a planned acquisition from a number and profile photo that differed from CEO Benedetto Vigna’s. A follow-up call mimicked Vigna’s southern Italian accent and pitched a “China-related deal” requiring a currency-hedge transaction. The executive asked which book Vigna had recently recommended; the caller hung up. Ferrari opened an internal investigation.
UK energy firm, March 2019. The British CEO of a UK energy firm took a call the executive believed came from the chief executive of the German parent company and, on an urgent demand, wired €220,000 (about US$243,000) to a Hungarian supplier. Sophos’s account, drawing on insurer Euler Hermes, records two further calls: a false claim of reimbursement and a second payment request from an Austrian number, which was refused. Sophos noted that AI involvement was the analysts’ belief, not a confirmed finding.
How do deepfakes get past identity checks?
Deepfakes defeat identity checks by substituting synthetic or swapped faces, voices and documents for the real person during onboarding or a video call. FinCEN’s Alert FIN-2024-Alert004, issued November 13, 2024, followed a rise in suspicious activity reports from 2023 into 2024 describing suspected deepfake media, including fraudulent identity documents used to circumvent verification.
Hong Kong’s face-swap loan syndicate is the case-level illustration. Between September 2022 and July 2023, the group allegedly stole identities and used an AI face-swapping program to apply for online loans from finance companies, totalling HK$200,000. Police arrested nine people, including the mastermind, in August 2023 on suspicion of conspiracy to defraud, according to the same LegCo reply.
FinCEN’s nine red flags give onboarding teams concrete signals. Among them: a customer photo that conflicts with the stated date of birth, a third-party webcam plugin during live verification, refusal of multifactor authentication, a reverse-image match to a gallery of AI-generated faces, and a new account sending rapid payments to gambling sites or digital-asset exchanges. FinCEN asks filers to put the key term FIN-2024-DEEPFAKEFRAUD in SAR field 2 and the narrative.
The same tooling works on consumers. In October 2024, Hong Kong police arrested 27 people in a romance-scam syndicate that, per The Record, sent victims AI-generated photos of attractive personas, then used face-swapping on video calls to alter the scammers’ appearances and voices. Losses reached about US$46 million across victims in Hong Kong, mainland China, Taiwan, India and Singapore. The FBI’s December 2024 PSA on generative AI fraud lists the same move: sending video to “prove” an online contact is a real person.
Why do attackers clone the voices of senior officials?
Account access, according to the FBI. Its May 15, 2025 PSA (I-051525-PSA) reported that since April 2025, actors sent texts and AI-generated voice messages claiming to come from senior US officials, built rapport, and then sent malicious links under the guise of moving to a separate messaging platform. Compromised accounts can then be used to target the victim’s contacts.
The December 19, 2025 update (I-121925-PSA) pushed the timeline back to at least 2023 and widened the impersonated set to White House and Cabinet-level officials, members of Congress and senior state officials, with targets including family members and acquaintances. Conversations move quickly to Signal, Telegram or WhatsApp. Requests include authentication codes that sync contacts, passport details, overseas wires and introductions. Once a contact list is in hand, the next round impersonates the previous victim.
Analysis: what patterns repeat across deepfake fraud cases?
Four patterns recur across the eight cases: public footage as raw material, channel switching, secrecy and urgency, and human rather than technical detection of the attempts that failed. None depends on a novel model capability.
Public footage is the raw material. Hong Kong police tied both 2024 video-conference cases to public online clips; the WPP attempt used YouTube footage. Executives who speak publicly supply the training data.
The attacker changes channels. Email to video (Arup), WhatsApp to Teams (WPP), WhatsApp to phone (Ferrari), text to encrypted messaging (the FBI campaign). The switch itself is a signal, and FinCEN flags the onboarding equivalent.
Secrecy and urgency replace verification. Confidential transactions at Arup, an acquisition and currency hedge at Ferrari, an urgent wire in 2019. Each framing discourages the victim from checking with a colleague.
Process caught the failures. Detection has a role, covered in the deepfake detection tools review, but it sits behind verification steps that do not depend on spotting the fake.
What should defenders do?
Close the callback gap first, then train the people attackers target.
- Within 7 days, brief executive assistants, board members and government-affairs staff on the FBI’s senior-official pattern: unsolicited voice messages, pressure to move to Signal, Telegram or WhatsApp, and any request for an authentication code.
- Within 30 days, require a callback through a directory number on file for any payment, bank-detail change or hedging instruction that originates in a video call, voice call or chat, regardless of the requester’s seniority or claimed confidentiality.
- Within 30 days, give executives and finance approvers a shared challenge phrase or question set. The FBI recommends a secret word for families; Ferrari’s executive improvised the corporate equivalent.
- Within 60 days, for regulated financial institutions, map FinCEN’s nine red flags into onboarding rules, flag virtual webcam plugins in liveness sessions, and train filers on the FIN-2024-DEEPFAKEFRAUD key term.
- Within 90 days, add a deepfake-impersonation scenario to the AI incident-response playbook, with bank recall contacts and IC3 reporting steps pre-filled, and confirm which notification obligations a loss would trigger.
FAQ
how can you tell if a video call is a deepfake
Test the caller with movements and questions a recording cannot fake. Hong Kong police advise asking the person on camera to turn their head up, down, left or right while watching for image errors, and probing with questions only the real person could answer. Pre-recorded fakes like the Arup call cannot respond, so verify payment requests by calling back.
is deepfake fraud illegal
Yes. In the cases on record, police treat deepfake fraud as ordinary fraud; the synthetic media is the method, not a loophole. Hong Kong police arrested nine people in August 2023 on suspicion of conspiracy to defraud for using AI face-swapping to obtain online loans, and in October 2024 arrested 27 people tied to a deepfake romance-scam syndicate.
where do i report deepfake fraud
In the US, report it to the FBI’s Internet Crime Complaint Center at ic3.gov, the channel both 2025 FBI advisories name, and call your bank immediately to request a recall. US financial institutions filing suspicious activity reports should include FinCEN’s key term FIN-2024-DEEPFAKEFRAUD. In Hong Kong, the police Anti Scam Helpline is 18222.
do deepfake detection tools stop these scams
Not on their own. In the three stopped attempts, at WPP, Ferrari and the 2019 UK energy firm, reporting credits no detection software; Ferrari’s turned on a verification question, the UK case on a mismatched number and a missing reimbursement. Detection adds a second layer, but a mandatory callback on payment requests closes the gap the Arup attackers used.
Related across the network
- Deepfake Cybersecurity: Attack Vectors and Defenses for 2026 — ai-alert.org
- Dutch Police Arrest ShinyHunters Suspect; Attacks Escalate — ai-alert.org
