The CEO on the video call was not real: deepfake fraud 2024-2026

Series: How minds are moved · Social engineering: the attack on one mind at a time · 3 of 6

In January 2024, a finance employee in Arup’s Hong Kong office joined a video call with the company’s chief financial officer and several colleagues. They looked right and they sounded right. Over about a week, following their instructions, he sent US$25.6 million (HK$200 million) to five bank accounts in 15 transfers. None of the people on the call were real. Arup later confirmed that “fake voices and images” had been used.

That case made the headlines. This post looks at the quieter pattern behind it: a cloned voice or face used to get money from a company or a family, and the simple rules that beat it.

Contents · 9 sections

TL;DR #

• Deepfake fraud uses AI-generated voices or video to impersonate someone the victim trusts: a CFO, a minister, a grandchild.
• The documented cases include Arup (lost about US$25.6 million), WPP and Ferrari (both attempts failed) and a 2025 Italian scam that used the defence minister’s cloned voice.
• In its 2025 report the FBI logged 22,364 AI-related complaints and $893 million in losses, the first year it tracked AI.
• The fake is the new part. The scam underneath is old: authority, urgency and secrecy.
• What defeats it costs nothing: a family code word, a call-back to a known number, and two people approving any unusual payment.

What changed, and what didn’t #

Impersonation fraud is old. “CEO fraud”, where someone pretends to be the boss and orders an urgent transfer, has been around as long as email. What generative AI changed is the cost of playing the role. A convincing voice used to require a skilled mimic. Now it takes a few public clips of the person speaking, a commercial tool and an afternoon.

The FBI said so plainly in a public service announcement in December 2024: generative AI “reduces the time and effort criminals must expend to deceive their targets”. The same announcement gave advice that sounds almost old-fashioned. Agree on a secret word or phrase with your family, so that you can check who is really on the phone.

I’m keeping this post on fraud against people and companies. Political deepfakes and election manipulation are a different family in this series.

Documented cases #

Arup, Hong Kong, January 2024. Hong Kong police described the case in February without naming the company, and Arup confirmed it in May. The employee had first been suspicious of an email about a “secret transaction”. The video call removed his doubts, because the people on screen looked and sounded like colleagues he knew. It was one of the first widely reported cases where a whole group video call, not just a voice, was faked.

WPP, May 2024. Fraudsters set up a WhatsApp account with the photo of WPP’s chief executive, Mark Read, and arranged a Microsoft Teams meeting using a cloned voice and video footage to impersonate him. The target was an agency leader, and the aim was money and personal details. It failed. Read later warned staff about it in an internal email, which the press reported.

Ferrari, July 2024. As Bloomberg reported, a Ferrari executive received WhatsApp messages from someone posing as the chief executive, Benedetto Vigna, about a confidential acquisition, followed by a call in a voice that copied his southern Italian accent. The executive noticed small oddities and asked a question that only the real Vigna could answer. The caller hung up. One question was enough.

Italy, February 2025. Callers using a cloned voice of the defence minister, Guido Crosetto, phoned some of Italy’s best-known business people. They said Italian journalists had been kidnapped abroad and that ransom money was needed urgently, to be repaid later by the Bank of Italy. Several targets checked and refused. One of Armani’s staff asked for a formal written request from the ministry, which never came. Italian media reported that at least one entrepreneur, the former Inter Milan owner Massimo Moratti, sent about €1 million before realising, and that police moved to freeze the money. Crosetto said he would file a complaint, and prosecutors in Milan opened an investigation.

“Grandparent” scams. In early 2023, police in St. John’s, Newfoundland, said at least eight seniors had lost a combined 200,000 Canadian dollars in three days to callers who claimed a grandchild was in jail and needed bail. One victim told the CBC he was sure the voice was his granddaughter’s. The CBC reported that AI voice cloning was likely, but police did not confirm how the voices were made. Canadian and US police have kept warning about the same script since.

How the scam is built, seen from the receiving end #

I’m not describing how to make a deepfake. I’m describing what the victim sees, because the shape is always the same.

1. A trusted figure appears in an unusual channel: a new WhatsApp number, a private video link, a call from a number you don’t know.
2. There is something to keep quiet, such as a confidential acquisition, a kidnapping, or a grandchild who “doesn’t want Mum to know”.
3. There is a deadline: today, before the court closes, before the market opens.
4. There is money, moved in a way that is hard to reverse: wire transfers, cryptocurrency, gift cards, cash collected by a courier.
5. The fake face or voice exists to remove one doubt, is this really them?, so that the authority, urgency and secrecy can do the rest.

That last point matters. The deepfake doesn’t replace the social engineering. It supports it.

What it feels like from the inside #

People describe it as a jolt of recognition followed by fear. It’s the boss’s voice, tense. It’s your grandson crying. Recognition switches off suspicion, and fear switches off thinking. A victim in the Newfoundland case said he was sure it was his granddaughter. The Arup employee’s early suspicion faded once he saw faces he knew.

Afterwards there is shame, and victims often hesitate to report, even to their own families. Police say that delay costs them the chance to recover money.

What deepfakes can and cannot do #

They can copy a voice from short public clips, and they can produce a passable face on a video call, especially with low resolution, a poor connection and a short meeting.

They are weaker at things that require real knowledge or real interaction: a private memory, an unexpected question, a request to turn the head or move a hand in front of the face, a call-back on a known number. The Ferrari attempt and most of the Crosetto calls failed for exactly these reasons.

Keep the scale in proportion as well. In the FBI’s 2025 figures, AI-related losses of $893 million were a small part of nearly $21 billion in reported online crime losses. Within investment fraud, losses tagged as AI-linked made up less than 8%. Most fraud is still done with plain text, a phone call and a good story. Deepfakes are the expensive edge of the same business, not a new business.

How to notice it #

• A familiar face or voice turns up on a new number, app or account.
• The request is secret, urgent and about money, all three at once.
• The person can’t or won’t switch channels: “don’t call my normal number, it’s monitored”.
• Small oddities: flat emotion, lag, a slightly wrong accent, a picture that doesn’t move naturally, background noise that doesn’t fit.
• You are asked to skip a normal approval step “just this once”.

Rules that work against any quality of fake:

1. Family code word. Agree on a word or phrase in person, never online. Anyone calling in an “emergency” must give it.
2. Call back. Hang up and call the person on the number you already have. For companies, use the directory, not the meeting invite.
3. Ask a question only the real person can answer, about something that is not on social media.
4. Two-person rule. No unusual payment is approved by one person on the strength of one call or one meeting, whoever seems to be asking.
5. Out-of-band confirmation for any change of bank details, every time.

Frequently asked questions #

How much audio does a voice clone need? #

Security vendors often say a few seconds. Quality improves with more audio and with a clean recording. Anyone with public videos, podcasts or voicemail greetings has enough material out there. The defence doesn’t depend on keeping your voice secret. It depends on verification.

Can detection software protect me? #

It can help companies screen calls, but it is an arms race, and families won’t run it on every phone call. A code word and a call-back work without any software.

What should I do if I already paid? #

Call your bank immediately, using the number on your card. Report to the police and, in the US, to the FBI’s IC3. Speed matters, because transfers can sometimes be stopped in the first hours.

Is it rude to ask my boss or my parent for a code word? #

No. The FBI recommends it. Agree on it calmly in advance, and it stops being an accusation and becomes a routine.

The lesson #

Deepfakes make the costume cheaper, but the play is the same one con artists have run for centuries: a trusted voice, a secret, a deadline, money that can’t come back. You don’t need to spot a perfect fake. You need habits that don’t depend on spotting anything at all: a word only your family knows, a number you dial yourself, and a second person who has to say yes. Arup’s employee met a convincing fake without those habits. The Ferrari executive met a good one with a single question, and that was enough.