The Outlier Lab

Case Study - Arup Deepfake Video Call (2024)

A Hong Kong finance employee at Arup sent HK$200M (about US$25.6M) in 15 transfers after a video call where every colleague was a deepfake.

What happened

  • Lure. The employee received a message that appeared to come from Arup's UK-based CFO about a confidential transaction. CNN reported that his first reaction was to suspect phishing.
  • Social proof. He was then pulled into a multi-person video conference. Police said everyone on the call except the victim was fabricated, built from publicly available video and audio of the people being impersonated. They described the deepfake segments as pre-recorded with little real interaction. Seeing and hearing several familiar colleagues is what overcame his doubts.
  • Execution. He made 15 transfers to five Hong Kong bank accounts (FT reporting via Dezeen).
  • Discovery. The fraud came to light only when the employee later checked with head office. Police received the report on January 29, 2024, after the money had already moved.
  • Scope. Arup said fake voices and images were used, that none of its internal systems were compromised, and that its financial stability was not affected. Arup's global CIO, Rob Greig, later called it "technology-enhanced social engineering."

The loss happened when a legitimate, authenticated user authorized a legitimate payment. Login, device and credential controls played no part. This is an APP scam with a corporate victim. Corporate AP fraud is otherwise out of scope for this vault, but the mechanics match the voice-clone impersonation scams consumers now face. The receiving side is squarely a bank transaction-monitoring problem: five mule accounts took in about US$25 million.

Diagram: Message from 'CFO' (confidential deal), Multi-person deepfake video call, Employee authorizes 15 transfers, 5 local bank accounts (mules), Funds dispersed, Later check with head office, Police report Jan 29, 2024

Timeline

DateEvent
January 2024Hong Kong finance employee gets the "CFO" message, joins the deepfake video call, and makes 15 transfers totaling HK$200M
January 29, 2024Hong Kong police receive the report, after all transfers had completed
Early February 2024Acting Senior Superintendent Baron Chan describes the case publicly without naming the company. No arrests.
February 2024At the same briefing, police describe a separate pattern: deepfakes used at least 20 times to beat facial recognition with stolen ID cards; 8 stolen cards tied to 54 bank account registrations and 90 loan applications (July–September 2023); 6 arrests in related scams
May 2024Arup is identified as the victim (FT, CNN, Dezeen) and confirms fake voices and images were used
June 26, 2024Secretary for Security Chris Tang tells LegCo that Hong Kong has had three deepfake video scams since 2023: an AI face-swap loan-fraud syndicate (HK$200K, nine arrests in August 2023), this HK$200M pre-recorded video-conference case, and a May 2024 "fake boss" video call seeking nearly HK$4M. He reports arrests only in the first case.
February 2025Greig discusses the lessons with the World Economic Forum. He says he built a crude real-time deepfake of himself with open-source software in about 45 minutes.

Mechanisms involved

Control gaps

Public sources don't describe Arup's internal payment approvals. The points below are inferred from the reported facts.

GapWhat the facts show
Video treated as verificationOut-of-band confirmation with head office happened only after the transfers
Confidentiality overriding processThe "confidential transaction" framing cut the approver off from colleagues who could have challenged it
New-beneficiary exposureLarge sums went to five beneficiaries with no apparent payment history
Splitting15 transfers. Public sources don't say whether any approval thresholds applied.
Receiving-bank monitoringFive accounts could take in large corporate inflows and pay them out before anyone intervened

Detection and response

No control detected the fraud; the victim's own follow-up did. By the time police were involved, the funds had already moved through the receiving accounts. Greig's account of Arup's response stresses three things: quickly establishing scope (no systems compromised, no data affected), being open about the attack internally and externally, and rehearsing responses to classes of incident rather than specific ones.

Outcome

  • The loss was HK$200M (about US$25.6M, or £20M). Arup said its operations were unaffected.
  • Police said the investigation was ongoing and reported no arrests. Later press (May 2024) also reported none, and the security secretary's June 2024 LegCo statement listed no arrests for this case. As of October 2026, no public report of arrests, prosecutions or recovered funds was found.
  • The case became the standard reference example for deepfake-enabled payment fraud in executive briefings and regulator speeches.

Lessons for practitioners

  1. Authenticating the person doesn't authorize the payment. Verify the payment instruction itself, through a channel the requester did not supply (a callback to a number on file, or a second approver who was not on the call). Process controls like these hold up as deepfakes improve; "spot the fake" detection does not.
  2. Treat audio and video as unauthenticated channels. Write it into policy: no payment is released on the strength of a call or video alone, however senior the requester. The consumer version is the same. Scam intervention scripts should assume the victim has already "verified" the request by phone. See Playbook - Scam Intervention Members and Scam Victim Indicators.
  3. Encode the scam triad as a hard stop. Secrecy, urgency and a new beneficiary together should force friction instead of only raising a score. At a bank this looks like Payee Confirmation and Payment Friction, Payee and Counterparty Risk and Velocity and Exposure Limits on outbound business payments split across new payees.
  4. The receiving side is the most detectable part. Large, out-of-profile inflows that leave quickly are classic Fast In Fast Out Patterns. Five accounts absorbing about US$25M should have tripped Transaction Monitoring at the receiving banks. See Playbook - Money Mule Detection Members.
  5. Deepfakes hit onboarding too. The same briefing tied synthetic faces to opening accounts with stolen IDs. Liveness and face-match vendors should be tested against injected and synthetic media as well as printed photos. See Identity Verification Outcomes.
  6. Don't overweight novelty. Greig's point is that the deception is old and only the tooling is new. Budget for process redesign and red-team exercises (Threat Intelligence and Horizon Scanning, Playbook - Novel Attack Detection Members, Fraud Incident Response) before buying a deepfake detector.
IndustrySame pattern, different money path
Neobanks and Consumer FintechVoice-cloned relative or "bank fraud team" pushes an instant transfer
Brokerage and InvestingImpersonated advisor or executive requests a liquidation and wire
Payments and P2PBusiness payer releases a real-time payment after a convincing call

Sources