Anatomy of the Arup $25M Deepfake Heist: How a Fake CFO Video Call Cleared 15 Wire Transfers

I'm Ryuta Hamamoto from TIMEWELL.

A finance employee at a Hong Kong office receives an email that appears to come from the CFO at UK headquarters. It asks them to join a video conference about a confidential transaction. They suspect phishing at first. Then the call opens, and the CFO and some colleagues are on screen. The doubt fades, and the payments go out.

This is not a movie plot. It happened to the British engineering firm Arup in January 2024. According to Hong Kong police, the loss was about HK$200 million (roughly US$25.6 million). It became the reference case for AI-assisted fraud against companies, and it pushed finance and security leaders everywhere to look again at how they approve payments.

In this article I set out what has actually been made public, using primary sources where they exist. Press coverage and the police account differ on one important point, and I flag it. I then cover similar cases and official responses as of October 2026, and five controls you can deploy quickly. I am not criticizing the company or the employee. The same method can be aimed at any organization.

Three takeaways

• Seeing a face and hearing a voice on a screen is no longer enough to confirm who someone is. Hong Kong police believe the meeting in this case was a pre-recorded video built from publicly available footage and audio. It did not need advanced real-time technology to work.
• The raw material can come from material already public. The point is not to blame executives for being visible. It is to design approvals on the assumption that fakes can be made from public footage.
• The real defense is process, not technology. Remove "I confirmed it on a video call" as an approval basis, and require confirmation through a separate channel plus more than one approver.

What happened: the public record

I have split the facts by source: an official reply from the Hong Kong government, Arup's own statement, and press reports.

Item	Detail	Source
January 2024	A Hong Kong employee received a phishing email from someone impersonating the CFO of the UK head office and was invited to a group video conference about confidential transactions	Hong Kong government reply to the Legislative Council (26 June 2024)
The meeting	Several people appeared on screen. Police believe it was pre-recorded, assembled from public video clips and voice samples, with no real interaction with the employee	Same
Payments	The employee authorized transfers to five local bank accounts, for a loss of about HK$200 million. After the meeting ended, further payment instructions arrived by instant message	Same
Number of transfers	15 transfers totaling about HK$200 million (about US$25.6 million)	Press reports citing police
First reaction	The employee first suspected phishing but set the doubt aside after the video call	Press reports citing police
Reporting	Arup notified Hong Kong police in January 2024	Press reports (Arup)
May 2024	Arup confirmed it was the victim and said fake voices and images were used. It said its financial stability and operations were unaffected and that none of its internal systems were compromised	Press reports (Arup statement)
Investigation	Still under investigation as of June 2024. I could not find a public announcement of arrests or charges as of 29 October 2026	Same

One caution on reading all this. The case is often described as a live video meeting in which an uncanny copy of the CFO responded in real time. The Hong Kong government's reply says police believe it was pre-recorded. From public information alone I cannot say which description is closer to what happened. I use the police account as the baseline here.

Either way, the practical lesson is the same. People were on screen, the voices sounded right, and several of them seemed to take the request for granted. That combination was used to switch off the checking step.

Why this method can work: three structural reasons

Framing this as one employee's lapse leads to the wrong fixes. Here are three conditions that would apply to almost anyone.

1. Distant executives are not people you deal with every day

For a branch employee, a head-office executive is rarely someone they talk to regularly. A request coming from that direction is not unusual in a global company. Footage and audio of executives also often sit in public, in investor presentations, talks and videos. Hong Kong police believe publicly available video and voice samples were the raw material here.

So the more visible an executive is, the easier the material is to collect. That is a structural condition for many listed companies. The practical answer is not less visibility. It is to design approvals so they do not depend on seeing or hearing someone.

2. Several people on screen created a sense of safety

A one-on-one call might have given someone pause. Here, other people appeared alongside the CFO. When several people seem to proceed from the same request, it is hard to doubt it. The attackers exploit that.

3. "Confidential" switched off the normal checks

"It's confidential, so don't use the usual route" removes dual approval, line-manager sign-off and callbacks in one sentence. It is a classic social-engineering line, now backed by fabricated video and audio. The reply also notes that further instructions came by message after the meeting, so the points at which anyone could verify were spread across channels.

Similar cases: Arup is not an exception

Deepfake impersonation of executives has produced both losses and near misses since 2024. Here is what I can confirm, with the source of each.

• Ferrari (July 2024). A call impersonating the chief executive by voice failed after an executive asked for the title of a book the real CEO had recently recommended, and the caller could not answer, according to press reports.
• WPP (May 2024). An attempt using a fake messaging account and a voice clone in an online meeting to impersonate the chief executive. The company said the attempt was stopped thanks to the vigilance of its staff, according to press reports.
• Another Hong Kong case (May 2024). According to the Hong Kong government's reply, an employee lost nearly HK$4 million in a video conference with someone posing as the boss, and the case is under investigation. The same reply says police were aware of three deepfake-related cases up to the end of May 2024.
• An Italian banking group's private bank (2026). In late September 2026, press reports said about 95 million euros was sent to overseas accounts after fake messages and voice calls impersonating senior executives and a lawyer. The reports rely on unnamed sources, the companies declined to comment, and the facts and the state of the investigation are not settled. The reports say part of the money has been recovered.
• Japan (February 2026). The National Police Agency published a warning about "fake president" fraud. Emails posing as an executive steer staff into a social-media group, ask for account balances, then request a transfer. It is not a deepfake-specific warning, but it shares the same executive-impersonation structure.

An earlier version of this article quoted a year-on-year growth rate for detections and a multiple for deepfake cases in Japan. I could not confirm those figures against primary sources, so I have removed them. I would rather fix the process than use numbers to raise alarm.

Where things stand in October 2026 (then vs. now)

Here is what has moved since the incident in 2024 and since I first published in May 2026.

• The Hong Kong investigation. It was still open in June 2024. I found no public update since.
• Hong Kong authorities' response. Police say they share intelligence with Interpol, overseas agencies and the AI industry, have an advisory panel of technology experts, and run public awareness campaigns. For suspicious video calls, they advise the public to ask the person to move their head, test them with conversation, and confirm remittance requests by calling back. A June 2026 reply to the Legislative Council describes deepfake detection being built into facial-recognition steps in online identity checks and a plan to begin testing step-up authentication within 2026. These mainly concern individual identity verification, not corporate payment approval.
• United States. On 13 November 2024, the Treasury's Financial Crimes Enforcement Network (FinCEN) issued an alert on fraud schemes using deepfake media against financial institutions and their customers, with red flags and reporting guidance.
• Japan. The National Police Agency published its "fake president" warning on 13 February 2026, calling for internal information sharing, a review of payment rules and caution about being steered into social-media groups. Japan also published the AI Guidelines for Business, version 1.2 (METI and MIC, 31 March 2026), which continues the push for organizational responses to AI misuse. The guidelines are guidance for businesses, not a mechanism that stops an Arup-style fraudulent transfer. I see redesigning the payment process as the effective line of defense. How to carry a governance framework into the organization connects with the ideas in governed enterprise AI agents.
• What has not changed. The cases that were stopped (Ferrari, WPP) share a pattern. Appearance did not stop them. A question only the real person could answer, and an employee who noticed something was off, did.

Five controls you can deploy starting tomorrow

1. Remove "confirmed on a video call" from your approval criteria

This is the central one. Explicitly strike "confirmed with management on a video call" from the conditions for authorizing payments, contract signing or disclosure of sensitive information.

Replace it with:

• A callback through a separate channel (a corporate phone, an identity-verified internal chat), using contact details from your own directory rather than ones supplied in the email or chat
• Dual approval recorded in your formal workflow system
• For transactions above a threshold, execution only on a later business day, not during the meeting itself

2. Add a shared secret or a question only the real person can answer

Between executives and finance leads, agree on a phrase or a question that is never written down. Treat this as a supplement. Rotate it, and do not rely on it as the last line. The main defense is the approval flow in point 1.

3. Review how public footage relates to your approval steps

Investor videos and talks are part of doing business. The aim is not to stop publishing. It is to check that no approval step depends on someone's face or voice. Setting reuse conditions for footage in contracts with media and partners is a small extra safeguard.

4. Treat "skip the usual process because it's confidential" as a trigger to check more

Say this in training: a request to bypass the normal flow should make you check harder, not less. Design your process so confidential matters follow the normal flow plus confidentiality terms. Treat new instructions that arrive by message after a meeting the same way.

5. Keep deepfake detection tools as a supplement

Technology that judges the authenticity of video and audio is improving, but nothing offers complete detection today. If you adopt a tool, use it as an extra layer combined with shared secrets and out-of-band confirmation. Getting steps 1 to 4 right comes before choosing a product.

Two homework items for executives starting today

1. With the CFO and CISO, audit the last 30 days of payments and list every case where a video meeting or verbal instruction was the basis for approval
2. Put "redesigning identity verification and approvals for the deepfake era" on the next board agenda

This is not an IT problem. It is a management problem.

How WARP SECURITY treats this

TIMEWELL's WARP SECURITY treats the Arup-style deepfake BEC as one of five simulated incident exercises.

In the Executives course, leaders role-play "If this happened to us, who actually stops it?" Participants experience the cognitive biases at play from the seats of the finance officer, the line manager and the executive.

In the Practitioners course, we move into detection technologies, payment system design reviews and SOC alert design.

Summary

• The Arup case showed that "I saw them and heard them" is not enough to confirm identity. Hong Kong police believe the meeting was pre-recorded, so it worked without advanced real-time technology
• The weak point is where payment processes meet human cognitive biases. Fix it with process design, not with individual vigilance
• Defense is a three-piece set: take video confirmation out of approvals, require separate-channel confirmation with more than one approver, and change executive mindset
• Similar cases kept being reported into 2026. I have said where something could not be confirmed, and I have avoided using numbers to alarm

If your first instinct in "deepfake defense" is to evaluate detection tools, you have started in the wrong place. Process first. Tools second. Otherwise, the wire transfer button gets pressed.

References

• Arup confirms $25 million deepfake scam loss in Hong Kong - CNN (Arup statement and police quotes, 16 May 2024)
• LCQ9: Combating frauds involving deepfake - Hong Kong Government (Legislative Council reply) (26 June 2024)
• LCQ13: Measures to tackle identity theft of members of public - Hong Kong Government (Legislative Council reply) (3 June 2026)
• FinCEN Issues Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions - FinCEN (13 November 2024)
• Warning on fraud targeting corporations ("fake president" fraud) - National Police Agency of Japan (13 February 2026, in Japanese)
• Ferrari Deepfake Attempt - Fortune (27 July 2024)
• Deepfake scam targets CEO of world's biggest ad firm - The Guardian (11 May 2024)
• AI Guidelines for Business, version 1.2 - METI and MIC (31 March 2026, in Japanese)
• How deepfakes are being used in cybercrime - WEF (February 2025)