If you think deepfake CEO fraud prevention for small business is a problem for later, start with what happened to a company that had every resource you do not.

An employee at Arup's Hong Kong office joined a video call with his chief financial officer and several colleagues he recognised. His chief financial officer gave him instructions. He made fifteen transfers totalling roughly 25 million US dollars.

The CFO was never on that call. Neither was anyone else. He had spent the meeting talking to AI-generated video.

Arup is a British engineering firm with 18,000 employees and over 2 billion pounds in revenue - a company that can afford the best security money buys. Here is what they said afterwards:

"Our internal systems were not compromised."

And that is true. No one breached any firewalls. No server, no VPN password, no control of any kind. The criminals did not break a system. They broke a person - specifically, his trust in his own eyes and ears.

The money remained unrecovered. The police did not arrest anyone. Arup's head for East Asia left the role a few weeks later.

Arup has 18,000 employees, over 2 billion pounds in revenue, and enterprise-grade security. The tooling that beat all of it is now cheap enough that the same attack scales down to a twelve-person company with one bookkeeper. If you want more details they can be found here.

The part of that story that matters most

The detail everyone skips is this: the employee was suspicious.

He received a message from the London CFO about a confidential transaction and thought it might be phishing. His instinct fired correctly. He was doing exactly what security training tells people to do.

Then came the video conference invitation. He joined, saw a face he knew, heard a voice he knew, and recognised the colleagues sitting alongside.

The suspicion evaporated. Because if you can see it and hear it, surely it is real.

That is the whole attack. Not technical sophistication - the deliberate destruction of the one verification method humans instinctively trust most.

What a deepfake actually is

As plainly as possible: a fake produced by AI. Video, an image, or a voice that looks and sounds like a real person.

The model takes a sample. A few seconds of your voice from any recording, a handful of photos, and it can say anything in your voice. Or place your face onto someone else's video.

Here is the uncomfortable implication. If you have ever posted a clip online where you speak to camera - a conference talk, a webinar, a company video, a birthday message - the source material already exists. Nobody needs to hack you. It is enough that you are visible.

Three years ago we laughed at deepfakes. They were clumsy, the artefacts were obvious, and detection was easy. I have stopped laughing. The tools are now cheap, fast and genuinely good.

Three places you will meet one

1. A call or video "from the boss" or from family. The voice matches, so you do not question it. "Mum, I have had an accident, I need money" - in your child's voice. It is the oldest scam in the book running on new hardware.

2. Fake videos with celebrities. A well-known investor or presenter "recommending" a crypto platform. The face is real, the words are generated. You click, you deposit, the money disappears. That's it.

3. An attack on a company. A fake CEO on a video call asking finance for an urgent payment, live. Or simply calling the accountant directly. This is Business Email Compromise upgraded with a face.

The common denominator in all three is always the same: urgency and money. If a request has both, treat it as suspect regardless of who appears to be asking.

The technical signals, and why you should not rely on them

There are detection cues. I will give them to you, with an honest warning attached: they get less reliable every month.

• Edges of the face - watch whether hair and beard shimmer or blur at the boundary
• Blinking - deepfakes often blink strangely, or barely at all
• Lip sync - check whether mouth movement genuinely matches the audio
• Lighting - light on the face that does not match the rest of the frame

Learn them. Then understand the limitation: AI will keep getting better, and your eyes will not. Any defence built on visual inspection has an expiry date, and it is soon.

The defence that actually holds

Verification through a different channel. That is the whole strategy, and it does not degrade as the technology improves.

Boss calls with an unusual payment request? Hang up and call back on his known number - the one in your contacts, not the one that called you.

A family member asking for money? Call their number and confirm.

An urgent transfer request in a meeting? Confirm it on a second channel before it moves. Every time, including when it is inconvenient, including when you feel foolish.

And agree a safe word - with your family and inside your company. A phrase only you know. A real person gives it without hesitation. A deepfake cannot.

For a small business, this costs nothing to implement. Write down one rule: no payment above a set threshold gets executed on the strength of a single channel, regardless of who appears to be asking. That policy would have saved Arup 25 million dollars.

I have already encountered a few cases where the client received a request to make a payment to a new account. There are different tricks here, but the repose should always be to make a direct call to a company who asked as to do it. Do not trust an email or Teams call.

Why "we would notice" is not a plan

The most common response I get from business owners is some version of: my team would spot it.

Arup's employee spotted it. That is the uncomfortable part of the case. He identified the initial message as probably fraudulent, on instinct, with no training prompt required. Then the attackers presented evidence his brain treats as conclusive, and the correct suspicion collapsed.

You cannot train your way out of that. Human beings are built to treat direct sight and sound as final proof, and until about three years ago that was a reasonable heuristic. It is not one anymore, and no amount of awareness training rewrites the instinct.

Which is why the defence has to sit outside the individual. Not "be more suspicious", which does not scale and does not survive a convincing video, but a process that does not depend on anyone's judgement in the moment.

Deepfake CEO fraud prevention for small business: what to put in place this week

For a family, one conversation is enough. Agree a phrase. Explain to whoever needs to know that if a request for money ever arrives by phone or video, the phrase gets asked for, and that asking is never rude.

For a business, three things are worth writing down:

A payment verification rule. Above an agreed amount, no transfer executes on a single channel. The person authorising calls back on a number already in the system - never a number supplied in the request itself.

A named exception for urgency. This matters more than it sounds. Attackers exploit the fact that people bypass process when the boss is impatient. Make it explicit, in writing, that urgency is precisely when the check applies rather than when it is waived, and that nobody is penalised for slowing a payment down.

A safe word for executives. Anyone who can authorise money, and anyone who can be impersonated to authorise money, shares a phrase. It takes five minutes to agree and it defeats every voice and video clone currently available.

None of this requires a budget, a product, or a consultant. It requires a decision.

The demo that ends the argument

I once had a meeting with a client who was absolutely convinced this did not apply to him and CFO laugh on me that deepfake CEO fraud prevention for small business is not his case. Well, let's play...

I asked a friend who does this sort of work. In about fifteen minutes he cloned the client's voice from the client's own YouTube video. I played him a sentence he had never said.

The room went quiet. He said: "That sounds exactly like me."

That is the point. You do not have to be a celebrity for this to reach you. It is enough that you exist online, which in 2026 covers essentially everyone running a business.

Which is why verifying through a second channel is not paranoia. It is hygiene.

Bottom line

Do not trust a face or a voice. Trust verification.

Unusual request involving money? Hang up, call back on a number you already have, use the safe word. Every time.

The same verification rule defeats the lower-tech versions of this attack too - see how to avoid instant payment and text message scams. And if the attacker's entry point is your mailbox rather than a video call, the Microsoft 365 settings that stop it take five minutes.