Everyone on the video call was a deepfake except the Arup employee who wired HK$200 million

Fifteen transfers before anyone knew the request was fake

In late January 2024, a finance employee in the Hong Kong office of Arup, the London-based firm that engineered the Sydney Opera House and Beijing's national stadium, approved a series of bank transfers worth about HK$200 million, roughly US$25.6 million.¹ The money moved in fifteen transactions to five local accounts, and it was gone before the firm realized the instruction behind it was fraudulent.²¹

The request had come from what appeared to be the company's chief financial officer, based in the United Kingdom, and the employee approved it after joining a video conference with the CFO and several colleagues.¹ Every other person in that meeting was fabricated. As the police officer who later described the case put it, in this video conference "everyone you see is fake."³

Arup confirmed in May 2024 that it was the target, that one of its Hong Kong employees had paid out about US$25 million, and that "fake voices and images were used." Its own systems, the firm said, were never compromised.⁴

The message looked like a scam until the meeting reassured him

The approach began in the middle of January, with a message that appeared to come from the firm's UK-based CFO about a confidential, or "secret," transaction.³ The employee's first reaction was suspicion. The message had the shape of a phishing email, the kind staff are trained to distrust.⁵

What settled the doubt was the video conference. On the call, the CFO and the other colleagues looked and sounded like people the employee recognized, and after seeing them his early doubts fell away.⁵ He followed the instructions and authorized the fifteen transfers.

The fraud surfaced only afterward, when the employee checked the transaction with the firm's head office and learned that no such payment had been arranged.⁵ By the time Arup reported the case to Hong Kong police, on 29 January 2024, the HK$200 million had already left.²

The pre-recorded meeting the employee joined live

The version of this story that traveled fastest describes a live, interactive video call, with the fraudsters conversing in real time as the CFO and the staff. The official record is more specific.

The fullest account is the Hong Kong government's written reply to its Legislative Council, given in June 2024.¹ It states that the video conference was pre-recorded, assembled from public video and voice clips of the executive it impersonated, and that there was no real exchange between the employee and the criminals at all. Once the recording had delivered its instruction, the fraudster ended the meeting and moved the actual payment demands to instant messaging.¹

The police officer who briefed reporters, Acting Senior Superintendent Baron Chan Shun-ching, described it the same way: the deepfake videos were pre-recorded and did not involve dialogue with the victim.² So the employee never argued with a fake CFO or watched one answer him. He joined a real meeting and watched convincing recordings of people he knew, then took his orders by text.

The distance between the two versions is worth holding onto. A two-way, real-time video deepfake able to hold a live conversation would be a more advanced attack than the one the record describes, and this incident is not evidence that such a thing defeated Arup. What defeated Arup was simpler, and for that reason more available.

Why a familiar face stopped proving who was speaking

The check the employee used is the one nearly everyone uses. He recognized the faces and the voices, so he believed the people were who they appeared to be. That check works only as long as an appearance is hard to counterfeit, and that is the part that has changed.

Start with faces. A 2022 study in the journal PNAS found that people asked to tell AI-generated faces from real photographs did no better than chance, guessing right about 48 percent of the time, and rated the synthetic faces as slightly more trustworthy than the real ones.⁶ The study tested still images rather than live video, so it does not prove that a moving face on a call is undetectable. It does show that the raw material, a synthetic face a person cannot flag as fake, is already ordinary.

Voices are further along. Cloning a specific person's voice from a short clip is cheap and quick, as an earlier lesson on a faked Biden robocall showed. For an attacker, a recognizable voice is now a commodity rather than an obstacle.

Governments have watched this move from research into fraud. A 2023 information sheet from the NSA, FBI and CISA reports that tools which once took a professional days to weeks now produce convincing fakes with little or no skill, drawn from a market "flooded" with free, "plug-and-play" software, and that the most damaging attacks impersonate leaders and finance officers. It cites a 2019 case in which faked audio of an executive's voice was used to steal US$243,000 from a UK company.⁷ Europol, writing in 2022, noted that video streams can already be manipulated in real time and applied to videoconferencing, while treating a fully convincing real-time impersonation as an emerging capability rather than a finished one.⁸ That timing fits the Arup record: at the start of 2024, a pre-recorded fake was the dependable option.

A face and a voice on a call now prove only that someone can render that face and that voice. They no longer prove who is on the other end.

The defense checks the money, not the face

If recognizing the fraud is the goal, the defender is in a losing position. The fakes are improving, and the PNAS result shows that people cannot reliably pick a synthetic face out of a lineup even under ideal conditions.⁶ A control that depends on someone spotting a good forgery will fail against a better one.

The control that works does not try. Instead of authenticating the face, it authenticates the transaction, through a channel the attacker does not control. Call the executive back on a number you already hold, not one supplied in the request. Confirm large transfers with a second approver. Agree a code word in advance for urgent money movements. The same NSA, FBI and CISA sheet recommends exactly this: verification procedures for real-time communications, tests for "liveness," and mandatory multi-factor authentication, which it calls especially important for the procedures that release payments.⁷

The Arup employee was not careless. He did verify with head office, which is how the fraud came to light.⁵ The verification simply came after the money had left rather than before, and a confirmation that arrives too late protects nothing.

This is why the weakness is so widely shared. Any organization that lets a live video call or a familiar voice stand in for identity at the moment money moves is exposed, and that describes most of them. The fake does not have to be flawless or interactive. It only has to survive one meeting while the transfer clears. Arup employs 18,500 people across 34 offices and is not a careless firm,⁴ and a single convincing meeting was enough to move HK$200 million out of its accounts.