Insights/ Threat Landscape
Why your family and your business need a secret word
A Hong Kong finance worker paid $25 million to a video call of deepfaked executives. The FBI recommends a simple defence: a codeword an AI can't know.

In January 2024, a finance employee at the engineering firm Arup joined a video call with his CFO and several colleagues and, over fifteen transfers, sent them USD 25.6 million. None of them were real. Every face and every voice on that call was AI-generated, built, police believe, from publicly available video and audio of real Arup staff. The same technology now targets families, and the fix works for a boardroom and a kitchen table alike. It is a short, private phrase that only the real person could know, agreed before anyone needs it. It costs nothing, takes five minutes to set up, and an AI cannot answer it no matter how convincing it sounds.
Data sourcesFBI Internet Crime Complaint Center (2024 Annual Report) · Federal Trade Commission, Consumer Alerts · CNN Business, Arup deepfake reporting · Hong Kong Police Force · Australian Cyber Security CentreFigures reflect the most recent full-year reporting available as of August 2026. Fraud figures are based on reported losses; actual totals are higher, since most victims never report.
01How the threat has evolved
The voice on the phone used to be proof
For most of business and family life, hearing someone's voice or seeing their face on a call was the verification. A written instruction could be forged. A voice could not: not convincingly, not in real time, not without effort far beyond most criminals.
That assumption is now wrong, and it broke quickly. McAfee researchers found that as little as three seconds of audio can produce a recognisable voice clone: a single voicemail, a video posted online, a moment of someone speaking in a public webinar. Consumer voice-cloning plans cost a few dollars a month, and open-source tools can run a real-time face swap on a well-equipped laptop.
3 seconds
Of audio was enough to produce a recognisable clone of someone's voice in McAfee's 2023 research.
15
Transfers made by Arup's finance worker before anyone checked with head office.
<$20/mo
Typical cost of a consumer voice-cloning plan: capability that was out of reach for most criminals a few years ago.
What has actually changed is not the scam, impersonating a distressed relative or an urgent executive is one of the oldest tricks in fraud, but the quality bar a criminal now has to clear. It used to take a skilled impressionist. It now takes a public LinkedIn profile.
02The current trends
The $25 million video call
The clearest illustration is also the most expensive one on record. In January 2024, a finance worker at the Hong Kong office of Arup, the engineering firm behind the Sydney Opera House and Beijing's Bird's Nest stadium, received an email from someone claiming to be the company's UK-based CFO, asking for a confidential transaction. He was sceptical, until a video call resolved his doubt: familiar faces, familiar voices, several colleagues he recognised.
Every person on that call was an AI-generated deepfake, Police believe they were built from publicly available video and audio of real Arup staff. Over fifteen transfers to five Hong Kong bank accounts, he sent HK$200 million, about USD 25.6 million. The fraud surfaced only when he later checked with the company's real head office. Arup confirmed the incident to CNN in May 2024; there has been no public report of the money being recovered.
Nobody on the call had been real. Every face he saw and every voice he heard were AI, built from footage and audio anyone could find online.
The same pattern runs at household scale, just with smaller numbers and more of them. The FTC describes it plainly: a call comes in, panicked, sounding exactly like a grandchild or a child. A car accident, an arrest, an urgent need for cash, often with a plea not to tell the rest of the family. The FBI's most recent annual data puts fraud losses reported by Americans over 60 at USD 4.9 billion, up 43% year on year.
Lost by Arup to a single deepfaked video call, across fifteen wire transfers.
CNN Business, May 2024Reported to the FBI in business email compromise losses in 2024 alone.
FBI IC3 2024 Annual ReportReported fraud losses among Americans over 60 in 2024, up 43% on the year before.
FBI IC3 2024 Annual ReportBusinesses and families are being hit by the same technology aimed at two different vulnerabilities: a company's trust in the chain of command, and a family's instinct to act first and ask questions later when someone they love sounds like they're in danger.
03Why the obvious defences don't hold
You can't out-look a deepfake
The instinct is to get better at spotting the fake: watch for unnatural blinking, listen for a flat cadence, check whether the lighting matches. That instinct is already running out of runway. The tells that worked in 2023 are mostly gone; the tools have improved faster than most people's ability to notice.
Caller ID is not a defence either. Spoofing a phone number is cheap and widely available, and carrier protections only partly block it. A criminal can make the call appear to come from a number already saved in your phone.
Answering a personal question doesn't fully solve it either, not if the question is guessable or has ever appeared online. A pet's name, a childhood street, a school mascot: these sit in data breaches, social media bios, and public records, and a well-resourced attacker can find them faster than most people expect.
What all of this points to is a defence that doesn't depend on detecting the fake at all. It depends on asking for something that was never public in the first place.
04The fix, for families
A word an AI cannot know
This is the FBI's current, official guidance: agree a secret word or short phrase with your family, before you need it. The FTC adds the second half: don't trust the voice, hang up and call the person back on a number you know is theirs. If a call comes in sounding like a relative in distress, the first response is to ask for the word. A cloned voice can say anything: it cannot know something that was never said where a scraper could find it.
Grandparents don't need the whole threat model. They need one instruction: ask for the word, then hang up and call back on the number you already have.
The FBI gave this advice in its December 2024 alert on AI-enabled fraud, and repeated it in its December 2025 alert on scam calls that pose as kidnappings. It is one of the few pieces of fraud advice that requires no technology, no subscription, and no ongoing vigilance: just an agreement made once, while everyone is calm.
05The fix, for businesses
The same idea, with a process behind it
A business can't rely on everyone remembering a shared phrase the way a family can. It needs the equivalent built into a process that survives urgency, seniority, and pressure to move fast. Four things do most of the work.
None of this requires new technology. It requires a policy that exists before the call comes in, and staff who have rehearsed following it under the kind of pressure a real attack applies.
06Stop. Check. Then act.
The habit that actually holds
Everything here reduces to the same instinct Mono trains for in Click or Flick: stop before you act on urgency, check through a channel the attacker doesn't control, then act. A codeword is that habit made concrete: a single, rehearsed question that works whether the caller is a stranger, a machine, or someone who sounds exactly like family.
It won't stop every attack. Nothing does. But it closes the gap that deepfakes are specifically built to exploit: the moment where a familiar voice replaces a moment's doubt with trust. A five-minute conversation, agreed while everyone is calm, is what stands between a family and a fake emergency call, and between a finance team and a boardroom that never existed.
·Key takeaways
What this means for your organisation
Three seconds of audio, freely available tools, and a real-time deepfake call runs on a consumer laptop.
USD 25.6 million lost to a video call where every participant except the victim was AI-generated.
Visual and audio tells fade faster than most people's ability to spot them. Don't rely on noticing.
Agreed in person, never online, never guessable. Ask for it; if it's missing, hang up and call back.
Callback verification, dual authorisation, and a codeword for high-risk requests: no exceptions for seniority.
The two levers attackers combine on purpose. A policy that flags both together catches most of the rest.


