For years, the classic CEO fraud was an email. A message that appeared to come from the chief executive asked someone in finance to make an urgent, confidential payment. Training taught staff to be suspicious of email and to pick up the phone. Now attackers can fake the phone call too, and the video meeting. Voice cloning and real-time video manipulation have turned a well-understood scam into one that can fool careful people.
This guide explains how deepfake CEO fraud works, what happened in the most widely reported case, and which payment controls, callback procedures, training practices and detection tools actually protect a finance team. The short version: you cannot rely on recognizing a face or a voice anymore, so your process has to do the verifying.
Why Deepfake CEO Fraud Matters Now
Payment fraud through impersonation was already expensive before deepfakes. The FBI’s Internet Crime Complaint Center (IC3) reported about $2.8 billion in business email compromise losses in 2024 from 21,442 complaints, and nearly $8.5 billion over 2022 to 2024, according to a summary published by Nacha. Deepfakes do not replace these schemes; they make them more convincing at the exact moment a skeptical employee tries to verify.
Regulators have noticed. In November 2024, FinCEN issued alert FIN-2024-Alert004 on fraud schemes using deepfake media, noting a rise in suspicious activity reports describing suspected deepfakes. In May 2025, the FBI warned that since April 2025 criminals had been sending AI-generated voice messages impersonating senior US officials to build trust before stealing account access. The same techniques work against any organization with a recognizable leader.
How Voice And Video Impersonation Fraud Works
Step 1: Research
Attackers study the company. Earnings calls, conference talks, podcasts, LinkedIn videos and webinars provide audio and video of executives. Org charts, press releases and social profiles reveal who works in finance, who approves payments and which deals or acquisitions are in progress. A few minutes of clear audio is often enough to build a usable voice clone.
Step 2: The Opening Contact
The first approach usually arrives through a normal channel: an email from a look-alike domain, a compromised executive mailbox, a WhatsApp or Signal message, or a text claiming the executive has a new number. It sets up secrecy and urgency, often around a confidential acquisition, a regulatory matter or a supplier who must be paid today.
Step 3: The Deepfake Verification
When the target hesitates, the attacker offers what seems like proof: a voice call, a voice note or a video meeting. The cloned voice or face removes the doubt. Some attacks put several fake participants on the same call so the request looks like a group decision. The attacker controls the meeting and keeps it short, avoids open questions and pushes for action before the call ends.
Step 4: The Payment
The target sends money to new bank details, often split into several transfers to stay under approval limits or bank alerts. Funds move quickly through mule accounts. Recovery depends on reporting within hours, not days.
The Arup Case: What Was Publicly Reported
The best-known example involves Arup, the UK engineering firm. As reported by CNN in May 2024, a finance employee in the company’s Hong Kong office received a message that appeared to come from the UK office about a secret transaction. He initially suspected phishing. He then joined a video call with people he believed were the chief financial officer and other colleagues. They looked and sounded like people he recognized, but all of them were deepfake recreations. He went on to make 15 transfers totaling HK$200 million, about US$25.6 million.
Arup reported the matter to Hong Kong police in January 2024, and police described the case publicly in February. Arup confirmed to CNN that “fake voices and images were used,” and said its internal systems were not compromised and its financial stability was not affected. That last detail is the key lesson: nothing was hacked. The fraud worked entirely by persuading a person to follow a payment request that looked properly authorized.
Payment Verification Controls That Stop Deepfake Fraud
The strongest defense is a payment process that no single conversation can override, no matter who appears to be on the other end. These controls should be written down, approved by the CFO and CEO, and announced to staff so everyone knows that executives will follow them too.
Dual Authorization For Payments
Require two people to approve any payment above a set threshold, with the approvals recorded in your banking or payables system. The second approver must review the supporting documents independently, not just take the first person’s word. Watch for attempts to split payments under the threshold, and set a lower limit for payments to new beneficiaries.
Treat New And Changed Bank Details As High Risk
Most impersonation fraud ends with money going to an account the company has never paid before, or to “updated” details for a known supplier. Put a hold on first payments to new beneficiaries and on any change to existing vendor bank details until they pass the callback procedure below. Many banks offer payee name checks or account validation services; turn them on where available.
No Exceptions For Urgency Or Secrecy
Write it into policy that no executive will ever ask finance to skip the payment process, keep a payment secret from the normal approvers or pay through a new channel. When a request includes any of these, the request itself is the red flag. This removes the social pressure from the employee: following the process is not disobeying the boss, it is doing what the boss already agreed to.
Use Bank-Side Controls
Ask your bank about positive pay, beneficiary allowlists, transfer limits, and dual control inside the banking portal. These controls work even if an employee is fully convinced, because the bank will not release the money without the second step.
Callback Procedures That Actually Work
A callback only helps if it goes to a number the attacker does not control. Many policies fail because staff call back the number in the email signature or reply on the same messaging thread.
- Call back using a number from an internal directory, HR system or vendor master file that was recorded before the request arrived. Never use contact details supplied in the request.
- For executive requests, call the executive on their known mobile or desk line, or confirm through a second person such as their assistant or another officer.
- For vendor bank changes, call a known contact at the vendor and confirm the change, ideally with someone other than the person who sent it.
- Do not accept a video call or voice note set up by the requester as verification. The requester controls that channel.
- Consider verification phrases or challenge questions for senior staff who can authorize payments, changed regularly and never shared by email. The FBI’s May 2025 warning recommends a similar code word approach for families.
- Log every callback: who was called, at what number, when and what was confirmed.
Training Your Finance Team
Training should change behavior, not just raise awareness. Focus on the people who can move money or change payment details: accounts payable, treasury, payroll, executive assistants and anyone with banking portal access.
Show real examples, including audio samples of cloned voices, so staff understand how convincing they are. Teach the pattern rather than the technical tells: unusual channel, urgency, secrecy, new bank details and pressure to skip steps. Visual and audio glitches can help, and the FBI suggests listening for odd tone, word choice or lag, but good deepfakes may show none of these. Make it explicit that staff will be supported, not blamed, for pausing a request to verify it, even if it turns out to be real.
Test the training. Run simulated requests through email, text and phone that follow the deepfake pattern, and measure whether staff follow the callback procedure. Firms that specialize in social engineering penetration testing can run these exercises, including voice phishing (vishing) campaigns, under an agreed scope.
Reduce The Attack Surface
You cannot remove every recording of your executives, but you can make the opening contact harder. Lock down email with SPF, DKIM and DMARC at enforcement, flag external senders and look-alike domains, and use phishing-resistant multi-factor authentication on executive and finance mailboxes so they cannot be taken over and used to send real-looking requests. Our guide to enterprise email security covers these steps in detail. Limit how much detail about payment approvers and finance processes is published on company websites and social profiles.
Deepfake Detection Tools: Where They Help
Detection tools analyze audio or video for signs of synthesis and can run on calls, voicemails and uploaded media. They fall into a few groups: real-time analysis plugins for video conferencing platforms, voice analysis for contact centers and phone systems, and identity verification services with liveness checks. FinCEN’s alert lists deepfake detection software, reverse image searches and metadata analysis among the tools financial institutions can use.
Treat detection as an extra layer, not the main control. Detection models can miss new generation techniques, and a false “looks real” result can make an employee more confident in a fake. Your payment process should stop a fraudulent transfer even if the deepfake is perfect. When you evaluate tools, ask which platforms they integrate with, how they report confidence, how they handle recordings and personal data, and what independent testing supports their accuracy claims.
What To Do If A Payment Has Already Gone Out
- Call your bank’s fraud line immediately and ask them to recall the transfer and contact the receiving bank.
- Report to law enforcement. In the US, file with the FBI at ic3.gov; IC3 runs a Recovery Asset Team that works with banks to freeze funds when reports arrive quickly.
- Preserve evidence: emails, message threads, call logs, meeting links and recordings.
- Check whether any mailbox or account was compromised, and reset credentials if so.
- Notify your cyber insurer and follow your incident response plan.
Frequently Asked Questions
Can Employees Reliably Spot A Deepfake On A Call?
Not reliably. Some fakes show glitches, odd lighting or lag, but quality is improving fast. That is why the protection has to come from verification through a separate channel and from payment controls, not from judging how someone looks or sounds.
Are Small Companies Targeted Too?
Yes. Smaller companies often have fewer approval layers and staff who know the owner personally, which makes a cloned voice more persuasive. Voice cloning tools are cheap, so the attack does not need a large target to be worth it.
What Is The Single Most Important Control?
A mandatory callback to a known number before paying any new beneficiary or changing bank details, combined with dual approval. Together, these stop most impersonation fraud regardless of how good the fake is.
Should Executives Stop Appearing In Videos And Podcasts?
Usually that is neither practical nor necessary. Public visibility has business value, and a determined attacker can often find enough audio anyway. Focus on making sure no voice or video, however convincing, can authorize a payment on its own.
Does Cyber Insurance Cover Deepfake Fraud?
It depends on the policy. Social engineering or funds transfer fraud coverage is often a separate endorsement with its own limits and conditions, and some policies require that verification procedures were followed. Review the wording with your broker before you need it.
Conclusion
Deepfake CEO fraud works because it attacks the moment of verification that older training relied on. The Arup case shows that a capable employee, on a video call with familiar faces, can still be deceived. The answer is a finance process built so that no single call, message or meeting can move money: dual approval, holds on new bank details, callbacks to numbers already on file, a written rule that urgency never overrides the process, and regular testing. Detection tools and training add useful layers, but the process is what keeps the money where it belongs.
