In early 2024, a finance employee at Arup — a multinational engineering firm behind projects like the Sydney Opera House — joined what looked like a routine video call with the company's UK-based CFO and several colleagues. He had already been skeptical of an earlier email asking him to arrange a confidential transaction. The video call is what removed his doubt.
Every face on that call was a deepfake, built from publicly available footage of real Arup executives, and every voice was synthetic. Over the course of the meeting the employee was talked into authorizing 15 separate wire transfers totaling roughly $25.6 million, sent to five different bank accounts in a single day. The fraud was only discovered when he later checked in with head office about the "confidential" deal — and was told no such meeting, no such transaction, had ever happened. As of early 2025 none of the money had been recovered.
Arup is not a small business, and $25 million is not a small-business number. But the mechanism that worked on its finance team is the same mechanism — scaled down — that is now being pointed at companies with five employees and one person who can move money. The technology that made the Arup call convincing is cheaper and more accessible than it was in 2024, not less.
What changed
- Fact. The Arup deepfake video call used AI-generated video and voice of real executives, built from public footage, to authorize $25.6 million in fraudulent transfers.
- Fact. Modern voice-cloning tools can produce a usable clone from roughly three to five seconds of clear audio — a voicemail greeting or a conference clip is enough.
- Fact. The FBI's Internet Crime Complaint Center logged 22,364 complaints carrying an AI-related descriptor in 2025, totaling $893 million in adjusted losses — and congressional researchers believe fewer than 5% of voice-clone scam victims ever report the loss.
- Fact. It isn't only Arup. WPP's own CEO, Mark Read, was impersonated by voice clone and old YouTube footage in an unsuccessful 2024 attempt to defraud one of his own agency leaders.
- Fact. The FTC's standard advice — call back on a number you already know is real, never one the caller gives you — still works. Almost nobody has a documented process for doing it under pressure.
01The scam is old. What changed is the disguise
"Urgent request from the boss, don't tell anyone else, move money now" is not a new scam. It is one of the oldest forms of business email compromise, and for years the defense against it was straightforward: written requests are easy to fake, so a real phone call from a familiar voice was treated as reasonably strong evidence.
That assumption is what deepfake voice and video attacks are built to exploit. The scam itself hasn't changed. What changed is that the thing you used to trust — recognizing a colleague's voice, seeing their face on a call — is no longer proof of anything. The Arup employee did the right instinctive thing when the email alone didn't convince him. He escalated to a call. The call is exactly what closed the gap.
02Why voice and video stopped being a reliable identity signal
Three things converged to make this practical against ordinary businesses, not just Fortune 500 finance departments:
- Training data got cheap to collect. A few seconds of someone's voice from a voicemail greeting, a webinar recording, a conference talk, or a LinkedIn video is enough to build a working clone. Executives, by the nature of their job, generate a lot of public audio and video.
- Real-time generation got good enough. Early voice-cloning demos were pre-recorded clips. The tools scammers use now can hold a live conversation — respond to questions, react in the moment — which is what made the Arup video call function as a real meeting instead of a suspicious monologue.
- Caller ID and email headers were never trustworthy in the first place. Spoofing a phone number or a sender name is old, cheap, and unrelated to AI. Deepfakes remove the next layer of doubt — the sound of the voice itself — that used to sit on top of those spoofed signals.
The three-minute version
If someone can hear a familiar voice on the phone or see a familiar face on a video call, that used to mean something. It no longer does, on its own. Treat any request to move money, change payment details, or approve an unusual transaction the same way regardless of how convincing the voice or video looks — verify through a separate, pre-agreed channel before acting, every time.
03What actually stops each version of this
The specific disguise varies. The verification step that defeats all of them is the same one: never trust the channel the request arrived on to also confirm the request.
| Scam pattern | How it's executed | What stops it |
|---|---|---|
| Cloned-voice phone call from "the boss" | A few seconds of public audio trains a clone; a call comes in urging an urgent, confidential wire | Hang up. Call back on a number already saved in your phone or CRM — never one the caller reads out. |
| Deepfake video call (the Arup pattern) | AI-generated video and voice of real executives on what looks like a normal meeting, requesting a transfer | A verified request never originates or gets approved inside the call itself. Confirm afterward, on a separate channel, before money moves. |
| Cloned "family member" or "colleague in trouble" call | Panic-inducing scenario, short window to act, request for money or gift cards | A pre-agreed code word the real person can say and a scammer cannot know. No code word, no action. |
| Compromised email plus a follow-up "confirming" call | A written request is reinforced by a phone call that sounds like the right person, to overcome hesitation | Dual approval on anything above a set threshold, with the second approver reached independently — not looped in by the first caller. |
Illustrative patterns based on documented 2024–2026 incidents, including the Arup case and the attempted WPP impersonation. Your specific controls should reflect your own approval chain and banking setup.
04Why "I would have known" is the wrong confidence
The Arup employee was not careless. He was skeptical enough of the initial email to want a call. The people on that call included several familiar colleagues, not just one impersonated executive, which is precisely what made it feel corroborated rather than suspicious. Mark Read's own agency leader at WPP was targeted the same year by a voice clone paired with real YouTube footage — a scam that failed, but only because something about the ask itself, not the voice, raised a flag.
The honest lesson from both incidents is not "be more skeptical." It's that skepticism aimed at the voice or the video is aimed at the wrong target now. The voice can be perfect. The face can be perfect. The only thing that reliably can't be faked is a callback to a number you already had before the call started.
05What to actually do
- Write down a verification protocol, don't rely on judgment in the moment. Under time pressure, from a familiar voice, in-the-moment judgment is exactly what deepfakes are designed to defeat.
- Callback to a known number, never a number the caller provides. Keep a short list of direct numbers for anyone who could plausibly request a wire — owner, partners, key vendors — stored somewhere the request itself can't redirect you to.
- Set up a code word for anything involving money movement. A phrase agreed on in person, changed periodically, that a real colleague can produce and a cloned voice cannot. The FBI recommends this for families; it works identically for a small business's approval chain.
- Require dual approval above a fixed dollar threshold, with the second approver contacted independently — not by the first caller "looping them in" on the same call.
- Never authorize a transfer, payment detail change, or urgent purchase from inside a single call or video meeting, however corroborated it feels. Confirm afterward, on a separate channel, before money moves.
- Treat "keep this confidential" as a red flag, not a compliment. Secrecy requests exist specifically to prevent you from checking with the one person who would say no.
- Tell your bank what your approval process is so a same-day, multi-transfer pattern like Arup's — fifteen wires in one day — has a chance of triggering a hold on their end too.
06The honest summary
The "urgent wire from the boss" scam is decades old. What's new is that the caller can now sound and look exactly like the boss, using nothing more than public audio and video that most executives already have plenty of online. That single change is why a firm the size of Arup lost $25.6 million to a fifteen-transfer scam in one day, and why the FBI logged over $893 million in AI-related fraud losses in 2025 alone — almost certainly a significant undercount.
None of that requires a technical defense you don't already understand. It requires a callback number you already had, a code word nobody outside the room knows, and a rule that no single call — however convincing — authorizes money to move on its own.
07Common questions
What actually happened in the Arup deepfake scam?
In early 2024, a finance employee at the engineering firm Arup joined a video call that appeared to include the company's UK-based CFO and several colleagues. Every face and voice on the call was an AI-generated deepfake built from public footage of real Arup executives. Over the course of the call the employee authorized 15 wire transfers totaling roughly $25.6 million, sent to five bank accounts in a single day. The fraud surfaced only when he later checked in with head office, which confirmed no such meeting or transaction had ever been authorized. As of early 2025 the money had not been recovered.
How much audio does it take to clone someone's voice?
As little as three to five seconds of clear speech is enough for modern voice-cloning tools to produce a usable clone, according to guidance repeated by both the FBI and FTC. A voicemail greeting, a clip from a webinar, or a short conference recording is more than sufficient — which matters because most business owners and executives already have far more than that circulating publicly.
Is this only a risk for large companies like Arup?
No. Arup is the best-documented case because of the size of the loss and Hong Kong police's public reporting, but the same mechanism scales down. A small business is arguably a softer target: fewer people in the approval chain, less formal verification process, and often a single person with authority to move money. The FBI's Internet Crime Complaint Center recorded over 22,000 complaints with an AI-related descriptor in 2025, and researchers believe most voice-clone scams go unreported entirely.
Why doesn't recognizing a colleague's voice or face on a call prove anything anymore?
Because both can now be synthesized convincingly and in real time — the attacker can respond to questions and react in the moment, not just play back a pre-recorded clip. That's what made the Arup call function as a believable meeting rather than a suspicious monologue. The verification value that a familiar voice or face used to carry has moved: it now belongs to a callback on a number you already had, not to the sound of the call itself.
What is a callback verification protocol and how do we set one up?
It means that any request to move money, change payment or banking details, or approve an unusual purchase gets confirmed by calling the requester back on a phone number you already had before the request arrived — never a number given to you during the call, and never by continuing the same call or video meeting. In practice: keep a short list of direct numbers for the owner, partners, and anyone who could plausibly request a wire, stored somewhere separate from the communication channel itself.
What is a code word and does it really help?
A code word is a phrase agreed on in person, in advance, that a real colleague or family member can produce on request and a scammer — even one using a perfect voice clone — cannot know. Both the FBI and FTC recommend it specifically for this reason: it doesn't depend on detecting anything about the voice, only on knowledge the clone doesn't have. Change it periodically and keep it out of email or chat where it could leak.
What threshold should trigger dual approval on a wire transfer?
There's no single number that fits every business, but the principle matters more than the figure: pick a dollar amount above which any transfer requires a second person's independent sign-off, reached on their own known number rather than being looped into the same call. Set the threshold low enough that it would have caught a transfer the size of your typical largest legitimate payment, not just an obviously large one.
Our bank already flags unusual activity — isn't that enough?
It helps, but it isn't a substitute for your own protocol, and the Arup case shows why: fifteen transfers in one day to five different accounts is exactly the kind of pattern a bank's fraud systems are built to catch, yet the transfers still went through. Telling your bank what your normal approval process and typical transfer patterns look like can improve the odds they flag a deviation — but the callback and dual-approval steps are the controls you actually control.
Write the callback list down this week
You don't need new software to close this gap — you need a short list of known-good phone numbers, a code word, and a rule that no single call authorizes a transfer. If you want a second set of eyes on your approval process for wires and payment changes, send us how it works today and we'll tell you the one or two steps most likely to matter. We don't sell fraud-prevention software, so there's nothing here for us to sell you.
Ask for a process checkSources, read 18 September 2026: Hong Kong police and Arup reporting on the 2024 deepfake video-call fraud that cost the firm approximately $25.6 million across 15 wire transfers (CNN Business, Fortune, and incident documentation from incidentdatabase.ai); reporting on the 2024 attempted deepfake voice-clone impersonation of WPP CEO Mark Read; the FBI Internet Crime Complaint Center's PSA on generative AI-facilitated fraud (ic3.gov/PSA/2024/PSA241203) and 2025 complaint data showing $893 million in AI-related adjusted losses; and FTC consumer guidance on voice-cloning scams and callback verification (consumer.ftc.gov). Related: The FBI Started Counting AI Fraud and An AI Agent Breached a Real Company.
Hero image from Unsplash, used under the Unsplash License.