Someone in your finance team takes a call from a voice that sounds exactly like your managing director, or joins a video call where the finance director and two colleagues are on screen. There is an acquisition, it is confidential, and a payment has to go today. Because the voices and faces have been generated to match people your team knows, everything about the request feels real.
What happened at Arup
In 2024 the engineering firm Arup confirmed it had lost 200 million Hong Kong dollars, about $25 million USD, to exactly this. Someone in its Hong Kong office received a message apparently from the UK-based chief financial officer about a confidential transaction, then joined a video call with what appeared to be the CFO and other colleagues. All of them were fakes.
That person made 15 transfers before checking with head office, which is when the fraud came to light. Arup said none of its systems had been compromised.
With no systems broken into, the attack depended on a payment process that trusted a familiar face on a call, and that process is something a business can change.
Why spotting the fake is not the answer
Cloned voices and generated video improve every few months, and detection tools are not reliable enough to carry the weight of a payment decision. The requests are also built to stop people thinking clearly: urgent, confidential and from someone senior enough that questioning them feels awkward.
What to change in the payment process
- Call back on a known number. For any payment above a set amount, any new payee and any change to bank details, the person approving calls the requester back on a number taken from your own records, never one supplied in the request.
- Require two people. A second person approves, and that person was not on the call or the email thread that made the request.
- Never change bank details on a message alone. A supplier's new bank details are confirmed through a contact you already hold, and a new payee waits a set period before the first payment goes.
- Remove the urgency exception. The rule says that an urgent, confidential request from a senior person is the case that gets checked hardest, and the leadership team agrees in writing that they will never ask anyone to skip the process.
- Consider a shared phrase. Some businesses agree a verification phrase for payment requests from leadership, changed regularly and never written in email.
- Rehearse it once a year. A planned test request, from a voice the finance team knows, shows whether the process holds under pressure, and thanks the person who stops it.
Tighten the technology around it
Although the process carries most of the weight, a few technical controls make the attacker's job harder. Email authentication on your domain makes it harder to send mail that appears to come from you. Strong sign-in on email and finance systems stops an attacker using a real mailbox to support the fake call. Dual approval in your banking platform makes the two-person rule technical rather than voluntary, and limiting who can add payees narrows the target.
Who needs to know
Everyone who can move money or change supplier records needs the process and the reason for it, and so do the assistants who manage leadership calendars and calls, because they are often the first contact. Leaders need it too, since they are the people being impersonated and the ones whose agreement makes the process stick.
Where this fits
Spotting AI-assisted attacks belongs in the everyone level of AI literacy, and it is one of the six places set out in AI security for a business that uses AI tools. If money does go, the first hours look much like the incident route for client data: get the facts, contain it by calling your bank at once and record every decision.