When a fake video or voice of someone you trust asks you to do something.
3 min to watch3 min to read
Based on a real case: Arup's Hong Kong office, early 2024.1,3,4 The screens, names, emails and lines are made up.
Read the walkthrough as text
Without a call-back
Hong Kong, early 2024. You're a finance worker. An email arrives from your company's UK office.
It asks for a secret transaction. You suspect a scam email.
Next comes a video call. The CFO and colleagues you know are on it.
The "CFO" explains: it's urgent, and it has to stay secret.
The others look and sound just like your colleagues. Your doubts go away.
You open the bank page and enter the first transfer.
You press Send.
Then again, and again: 15 transfers to 5 bank accounts.
In total: HK$200 million, about US$25 million.
Later, you check with head office.
Head office never asked for any of it.
Police later said everyone on the call except the worker was fake, made from public videos.
This is a deepfake: an AI copy of a real person's face and voice. Here, it was used to ask for money.
That's a deepfake attack.AI copies of real colleagues, made from public videos, got a worker to send HK$200 million. It happened at the engineering firm Arup's Hong Kong office in early 2024.
With a call-back
Same email, same video call, same request. This time the company has two payment rules.
You enter the first transfer and press Send. The bank holds it.
Rule one: a big payment needs a second person to say yes.
Rule two: call the CFO back, on the number in the company directory. Not one from the email.
The real CFO picks up. There was never any payment request.
So the video call was fake. The transfer is cancelled, and not one dollar leaves the company.
Hang up. Call back on a number you already have.A deepfake can copy a face and a voice. It can't answer the real CFO's phone, and it can't make a second person say yes.
Keep going:
Why it matters, what it costs, and what to ask your team.
The real case
One meeting, HK$200 million.
Early in 2024, a finance worker in the Hong Kong office of the engineering firm Arup joined a video call with what looked like the CFO and colleagues. Everyone else on it was a deepfake: an AI copy of a real person's face and voice.1,3
HK$200Msent, about US$25 million
0transfers
0bank accounts in Hong Kong
The worker had suspected the first email was a scam. The familiar faces on the call made those doubts go away.1 The truth came out only after a check with head office.4
Wait. Nobody hacked anything? It was just… a meeting?
Why it worked
The fake controls the call. You control the call back.
Nothing was hacked. Arup said none of its internal systems were compromised.2 The attack went after a person and a payment habit.
Police said the scammers used public video and audio of the real people, found on YouTube.3 So seeing someone is no longer proof it's them.
Asking "is it really you?" on the same call doesn't help. Switch between the two:
The fix has a name: out-of-band verification. Confirm the request on a route the requester didn't choose.
The scam itself is old. The FBI calls it business email compromise, and counts fake requests by email, phone and video meeting in it.5 Deepfakes just give it a face.
Source: FBI IC3 2024 Annual Report.
About 21,000 reports and close to $3 billion lost, every year. A steady business, not a fad.
I can copy a face. I can't pick up the real CFO's phone.
Ask your team
Which habit would have stopped it?
Every step of the attack was a chance to stop it. Turn on a habit and watch where the money stops.
Turn on the habits your company has. The attack replays each time.
1An email asks for a secret transactionfrom "the UK office"
2A video call: the CFO and colleaguesevery face and voice is fake
315 transfers to 5 bank accountsHK$200 million in total
$The money leavestoo late to take back
The three steps follow the real case. Which habit stops which step is simplified.
None of these is new software. They're rules about money. So ask your team:
Can one person send a big payment alone?If yes, one convincing call is all an attacker needs.
How do we confirm a money request?The answer should be: on a route the requester didn't pick, like a number from the company directory.
Is it safe to say "let me call you back" to the boss?Make checking the rule, even for the CEO. People only do it when it can't get them in trouble.
Under the hood, step by step: build payment rules that a fake CFO can't talk past.
Build along
Build payment rules a fake CFO can't talk past, one step at a time
We start with the simplest payment code, run it on the same fake request, and fix it in small steps. Press Run it after each step and watch the result change.
Detection
A detector is a dial, not a lock.
Deepfake detectors give each call a score. You pick a line, and anything below it gets flagged. Drag the line until every fake is caught:
Made-up scores to show the idea. Each bar is how real a detector thinks the call looks.
To catch the best fake, you flag real people too. Set the line lower, and good fakes walk through. That's why detection can warn you, but the call-back has to decide.
Use it as an alarmA warning should trigger the call-back and a second approver.
Expect it to miss someDetectors learn from yesterday's fakes. New tools make new ones.
Expect false alarmsBad cameras and weak Wi-Fi make real people look fake.
Never let it replace the rulesIf one person can still pay alone, one missed fake costs everything.
2024 IC3 Annual Report· FBI Internet Crime Complaint Center, 2025. Business email compromise: definition, losses and complaints for 2022–2024.
The emails, video-call lines and screens are recreated from news reports and simplified. The "with a call-back" replay is made up to show the fix. In the build-along, the HK$200 million is split evenly across the 15 test transfers, and the limit is made up. The detector scores and which habit stops which step are made up to show the idea.