Deepfake fraud is CEO fraud, vishing or business email compromise carried out with an AI-cloned voice or face of someone the victim trusts. The request is the familiar one (a payment, credentials, an MFA code or sensitive data), and the deepfake is the step that removes the victim’s doubt.
The FBI says AI-generated content “has advanced to the point that it is often difficult to identify” (FBI), so security awareness training has to teach a verification step that holds however convincing the voice or face is. The seven deepfake fraud examples below, from 2023 to 2025, show how the attack runs and what stopped it when it failed.
How deepfake fraud works
A typical attack runs in five steps:
- Collect the voice and face. Conference talks, YouTube footage and profile photos are enough.
- Make first contact on a side channel. A WhatsApp message from a new number with the boss’s photo, or an SMS followed almost at once by a move to Signal, Telegram or WhatsApp (FBI).
- Set a confidential pretext. A secret acquisition or restructuring, often with an NDA and a fake lawyer on the thread (Singapore Police Force).
- Deliver the deepfake. A voice message, a live call or a video meeting with several fake participants.
- Make the ask. An urgent transfer, often split into several payments to accounts the criminals control, or credentials, an MFA code or passport copies. A larger second request often follows.
Why deepfake fraud works
Deepfake fraud leans on old pressure points (authority, secrecy and urgency) and delivers them in a familiar face and voice. A video call feels like the verification step, so it switches off the doubt that a strange message created. A Department of Homeland Security assessment quoted by FinCEN says deepfakes “do not need to be particularly advanced or believable to be effective” because of “people’s natural inclination to believe what they see” (FinCEN).
The FBI says AI-generated content “has advanced to the point that it is often difficult to identify” (FBI). Training should teach a verification procedure that holds however good the fakes get.
Signs of deepfake fraud
The first four signs come from the request and the caller’s behaviour, so they still apply when a call has no glitches at all.
- A business request from a new number or an unusual channel, such as WhatsApp.
- Secrecy: a “secret acquisition, transaction or payment that no one else knows about”, in the words of WPP’s CEO (The Guardian).
- Pressure to move to another app or to skip normal approvals.
- A caller who ends the call when asked a personal question.
- A voice that is slightly off: mechanical intonation, call lag, unusual word choice, or not quite how the person usually speaks.
- On video, lip movement that doesn’t match the audio, unnatural movement or odd shadows.
Deepfake fraud examples
Arup’s deepfake CFO and colleagues (January 2024, engineering).
A Hong Kong finance worker joined a video call with the “CFO” and several colleagues, all deepfakes, and sent HK$200 million (about US$25.6 million) in 15 transfers. He had first suspected the message was phishing; the call convinced him. The fraud surfaced only when he later checked with head office. Arup confirmed that “fake voices and images were used”. Lesson: a video call with familiar faces is not verification. (CNN)
WPP’s cloned CEO in a Teams meeting (2024, advertising).
A fake WhatsApp account with CEO Mark Read’s public photo set up a Teams meeting with a voice clone and YouTube footage of Read. An agency leader was asked to set up a new business to extract money and personal details. The attempt failed. Lesson, in Read’s words to staff: “Just because the account has my photo doesn’t mean it’s me.” (The Guardian)
Ferrari’s cloned CEO stopped by one question (2024, automotive).
An executive got WhatsApp messages “from” CEO Benedetto Vigna about a big acquisition, then a live call in a convincing southern Italian accent asking for a currency-hedge transaction. He noticed mechanical intonations and asked which book Vigna had just recommended to him, and the call ended. Lesson: a personal question is a quick identity check that public footage can’t prepare a clone for. (Fortune, citing Bloomberg)
A deepfake CEO on Zoom in Singapore (2025, unnamed multinational).
A finance director was contacted on WhatsApp by a fake CFO, then joined by a fake lawyer in a Zoom “restructuring” meeting where the CEO and others were deepfaked. He transferred more than US$499,000. A second request, for US$1.4 million, made him realise it was a scam. HSBC alerted police and the full amount was recovered. Lesson: reporting fast can get the money back. (Singapore Police Force)
Italy’s cloned defence minister (2025, business leaders).
Callers posing as defence ministry officials passed the phone to an AI-cloned Guido Crosetto, who asked business leaders for money to free kidnapped journalists, saying the government couldn’t be seen to be behind it. Former Inter Milan owner Massimo Moratti sent almost €1 million in two transfers before police traced and froze it in a Dutch account. Lesson: a famous voice that asks for money and secrecy gets verified like anyone else. (Reuters via Yahoo)
Wiz’s cloned-CEO voice message (2024, cloud security).
Dozens of employees received a voice message in CEO Assaf Rappaport’s cloned voice asking for their credentials. The clone was trained on a conference talk, where his voice differs from his everyday one, and staff noticed. Lesson: “it doesn’t sound quite like him” is reason enough to stop. (TechCrunch)
Retool’s fake IT call for an MFA code (2023, software).
An SMS posing as IT led one employee to a fake login page. The attacker then phoned as IT in what Retool describes as a deepfake of an employee’s voice. The caller knew the office floor plan and colleagues, and talked the employee into one more MFA code. With that code the attacker took over the accounts of 27 cloud customers. Lesson: IT never needs an employee’s MFA code, by phone or any other way. (Retool)
How to prevent deepfake fraud
Personal habits to prevent deepfake fraud
- Verify requests for money, credentials or codes through a known number or an established channel, never the one the request came in on, because a voice or face on a call is not proof of identity.
- Check any deal that must skip approvals or stay hidden from colleagues, because secrecy is a warning sign in itself.
- Ask a question only the real person could answer. Treat a caller who hangs up instead of answering as a fraudster.
- Report a payment or a shared code to the bank and to IT at once.
Security controls to prevent deepfake fraud
- Write the call-back rule into the payment procedure, so no request skips it, however senior the voice or face on the call.