The Call That Was Not Real
In early 2024, an employee in the Hong Kong office of Arup — the multinational engineering firm behind the Sydney Opera House, the Beijing National Stadium, and London's Crossrail — joined a video conference call with the company's Chief Financial Officer and several senior executives.
The CFO explained an urgent, confidential transaction that required immediate wire transfers. The employee had initial doubts. But the CFO was right there on screen, speaking naturally, answering questions. The other executives on the call confirmed the instructions. Everything appeared legitimate.
The employee authorized 15 separate transactions totaling HK$200 million — approximately $25 million USD — to five different bank accounts.
Every single person on that video call was a deepfake. The CFO was not real. The executives were not real. The voices were synthesized. The video was generated in real time. The entire call was an AI-fabricated performance designed to steal $25 million.
How the Deepfakes Were Created
The attackers used publicly available video footage of Arup's executives — investor presentations, conference appearances, media interviews — to train AI models capable of generating realistic video and audio in real time. The technology required to pull this off is no longer theoretical or experimental. Real-time face-swapping and voice synthesis tools are available as open-source software and commercial products.
The deepfakes were convincing enough to pass the scrutiny of a finance professional on a live video call. The attackers replicated not just faces and voices, but mannerisms, speech patterns, and the casual dynamics of a multi-person executive meeting.
This was not a crude phishing email. It was a sophisticated, real-time social engineering attack that exploited the fundamental human tendency to trust what we see and hear.
Arup Was Not the Only Target
Shortly after the Arup incident became public, it was revealed that WPP CEO Mark Read had been targeted in a similar scheme. Attackers created a deepfake of Read on a Microsoft Teams call, attempting to impersonate him to a senior WPP executive. The attempt was identified and thwarted before any money changed hands, but the sophistication of the attack alarmed the global business community.
These are not isolated cases. They are the visible tip of a rapidly growing iceberg. Global losses from deepfake-enabled fraud exceeded $200 million in Q1 2025 alone. The average loss per incident now exceeds $500,000, with the largest cases — like Arup — reaching into the tens of millions.
Why Traditional Security Failed
Arup had standard corporate security protocols in place. Multi-factor authentication. Approval chains for large transactions. Verification procedures for unusual requests. None of it mattered.
The deepfake attack bypassed every traditional security measure because it exploited the one channel that most organizations still trust implicitly: live video. When you can see someone's face and hear their voice on a video call, the human brain treats it as verified. We are hardwired to trust what we perceive directly.
Email-based fraud has been mitigated by training, filters, and authentication protocols. Phone-based fraud has been reduced by caller verification and callback procedures. But video-based fraud has no widely deployed countermeasure. Most organizations have no protocol for verifying that the person on a video call is actually the person they appear to be.
The Verification Gap in Enterprise Communications
The Arup attack exposed a critical blind spot in enterprise security: there is no standard way to verify identity on a video call. We verify email senders, authenticate website sessions, and validate digital signatures on documents. But when a CFO appears on a Zoom or Teams call, we simply trust that it is the real person.
This verification gap is now being exploited at scale. The technology to create real-time deepfakes is improving faster than the technology to detect them — unless organizations proactively deploy detection tools.
The Financial Impact Beyond the $25 Million
The direct financial loss was $25 million. But the total cost to Arup extended far beyond the stolen funds. The incident triggered internal investigations, enhanced security reviews, and reputational damage. It raised questions about the company's financial controls and became a case study in corporate vulnerability.
For other enterprises, the Arup case sent a clear message: any organization that relies on video conferencing for decision-making — which is to say, every major enterprise — is vulnerable to this exact attack. The question is not whether your organization will be targeted, but when.
Insurance coverage for deepfake fraud is still evolving. Many cyber insurance policies do not explicitly cover losses from AI-impersonation attacks. Organizations that suffer these losses may find themselves without recourse.
The Real-Time Detection Imperative
The Arup attack succeeded because there was no real-time detection layer on the video call. No system was analyzing the video feed for synthetic artifacts. No tool was verifying the audio against known voice profiles. No technology was checking whether the CFO's face had been generated by an AI model.
All of these detection capabilities exist today. AI-generated video has detectable artifacts — subtle inconsistencies in facial movements, lighting, reflections, and audio synchronization that human observers miss but detection algorithms catch. AI-synthesized voices have spectral patterns that differ from natural speech in ways that are invisible to the human ear but measurable by forensic audio analysis.
The problem is deployment. These detection capabilities need to be integrated into the communication tools that enterprises already use — not bolted on as an afterthought, but embedded as a real-time verification layer.
How DeFake Protects Against This
DeFake provides enterprise-grade deepfake detection for the media that matters most: video calls, voice communications, and the images and documents that drive business decisions. Our multi-modal analysis engine detects AI-generated video and synthesized audio in real time, flagging potential deepfakes before they can be used to authorize fraudulent transactions.
For enterprises, DeFake offers API-level integration that can analyze video feeds, uploaded recordings, and audio files for signs of AI generation. Our detection engine evaluates facial consistency, audio spectral patterns, lip-sync accuracy, and dozens of other forensic indicators that distinguish real media from synthetic media.
The $25 million Arup attack was preventable. Real-time deepfake detection on the video call would have flagged every participant as synthetic. The employee would have been alerted before authorizing a single transaction.