In February 2024, every person on a video call with an Arup finance employee was a deepfake — including the CFO. Before the call ended, $25 million had been wired to Hong Kong accounts. The employee had done exactly what security training recommended: verified the request visually. It made no difference.
That case was a warning. What has happened since is an escalation so severe that the financial industry now has a term for it: synthetic executive fraud. The technology that made that single $25 million theft possible has since become cheaper, faster, more accessible, and nearly undetectable by human perception alone.
The numbers are not subtle. Deepfake fraud drained $1.1 billion from U.S. corporate accounts in 2025 — triple the $360 million lost the year before. The volume of deepfakes online grew from 500,000 in 2023 to over 8 million in 2025. CEO fraud now targets at least 400 companies per day using deepfake technology. And according to security researchers, human detection rates for high-quality video deepfakes stand at just 24.5%.
Your people cannot see it coming. The question is whether your architecture can.
How the Attack Works
The anatomy of a CEO deepfake attack has become disturbingly routine. Attackers don't need special access to your systems or inside knowledge of your operations. They need what your executives have already made freely available: their voices, their faces, and their patterns of communication — published across LinkedIn, earnings calls, conference keynotes, podcasts, and company webinars.
Every public appearance your executive makes — every earnings call, every keynote, every media interview — is simultaneously building their personal brand and supplying adversaries with the training data needed to clone them. The visibility that makes a leader influential makes them a target.
Voice cloning tools available today require as little as three seconds of audio to generate a convincing replica with 85% accuracy. As researchers at Fortune reported in December 2025, voice cloning has crossed what they term the "indistinguishable threshold" — human listeners can no longer reliably tell the difference between a cloned voice and the authentic one. The perceptual tells that once gave away synthetic audio — unnatural rhythm, metallic undertone, mismatched breathing — have largely been engineered out.
Video deepfakes have followed the same trajectory. Earlier generation deepfakes could be identified by forensic artifacts around the eyes and jawline, temporal flickering between frames, and the "uncanny valley" quality that triggered instinctive unease in viewers. Current generation models maintain temporal consistency across full motion video, produce coherent facial movement under varied lighting conditions, and can operate in real time — meaning the deepfake on the other side of your video call is responding to you live, not playing back a pre-recorded clip.
The sophistication of the social engineering layer has kept pace with the technology. Attackers have learned that finance professionals who have been briefed on deepfake threats will attempt to verify unusual requests via video call. So they now proactively suggest the video call themselves — turning the verification step into a vector. The appearance of willingness to be verified creates exactly the false confidence that bypasses the employee's remaining skepticism.
The Cases That Defined the Threat
A finance employee at global engineering firm Arup received a message claiming to be from the company's UK-based CFO describing an urgent, confidential transaction. The employee, appropriately cautious, requested a video call to verify the request. On that call appeared not just the CFO — but multiple senior colleagues, all speaking naturally and responding in real time. Every participant was an AI-generated deepfake.
The employee authorized 15 transactions totaling $25 million to Hong Kong bank accounts. Arup's global CIO later described the attack as part of a sharp rise in the number and sophistication of incidents the firm was encountering. The case shattered the assumption that video calls are inherently trustworthy — an assumption that had become the default verification protocol for precisely this kind of fraud.
A senior executive at a UK energy company received a phone call from someone whose voice was indistinguishable from the company CEO — including the accent, consonant patterns, and conversational cadence. The caller requested an urgent wire transfer to a supplier. The executive complied without hesitation. Only afterward did they discover the voice was entirely synthetic. This case, occurring in 2019 when voice cloning technology was a fraction of its current capability, established the playbook that attackers have since refined and scaled.
A finance director received contact from someone posing as the company CFO requesting an urgent wire transfer for a confidential acquisition. When the finance director expressed hesitation, the attacker proactively suggested a video call — a deliberate tactic learned from the Arup case. The video call featured a convincing real-time deepfake of the CFO. The apparent willingness to be verified created sufficient false confidence to complete the transfer.
"The number of deepfakes increased from 500,000 in 2023 to over eight million in 2025. The attack is no longer rare. It is routine."
Why Existing Defenses Are Failing
The standard enterprise response to fraud risk has been awareness training — teaching employees to recognize suspicious requests, verify unusual instructions, and treat urgency as a red flag. That model was designed for a world where seeing and hearing a person meant you were actually seeing and hearing them. That world no longer exists.
The deepfake attack specifically weaponizes the verification behavior that training instills. Employees have been told to verify requests by calling back on a known number or requesting a video call. Attackers now anticipate this response and engineer the attack to survive it. The video call is the attack, not a defense against it.
Detection technology faces its own crisis. The market for AI deepfake detection tools is growing rapidly — but the effectiveness of these tools drops by 45 to 50 percent when deployed against real-world deepfakes outside controlled laboratory conditions. Human detection rates for high-quality video deepfakes stand at 24.5%. In a 2025 study by iProov, only 0.1% of participants correctly identified all fake and real media presented to them.
Security training tells employees to verify unusual requests via video call. Attackers have engineered attacks specifically to pass video verification. The behavior that training produces has become a step in the attack chain, not a defense against it.
And critically: 80% of companies currently have no established protocol or response plan for handling a deepfake-based attack. Organizations that have invested in fraud awareness training, dual-authorization financial controls, and phishing simulation programs have almost universally failed to extend those frameworks to cover the synthetic executive threat vector. The governance gap is as significant as the technical one.
What Actually Works
The meaningful defense against synthetic executive fraud is not better human detection — that battle is already lost at the perception layer. It is architectural: building verification systems that cannot be bypassed regardless of how convincing the impersonation is.
The organizations that will avoid becoming case studies are not those with the most suspicious employees. They are the ones that have made the financial authorization process architecturally resistant to social engineering — regardless of how convincing the impersonation is, and regardless of whether a video call was part of the verification attempt.
"Simply looking harder at pixels will no longer be adequate. The meaningful line of defense will shift away from human judgment entirely."
The Board Imperative
Synthetic executive fraud is not a fraud team problem or an IT problem. It is a governance problem — and the data makes that clear. Deepfake fraud losses tripled year-over-year in 2025. The trajectory toward $40 billion in generative AI fraud losses by 2027 is not a projection made under optimistic assumptions. It reflects current growth rates applied forward.
Every board member sitting on an audit or risk committee should be asking three questions of management right now. First: which of our executives have significant public media exposure, and do we have enhanced verification protocols in place for financial instructions originating from or attributed to them? Second: does our financial authorization architecture require out-of-band verification that cannot be satisfied by a video call alone? Third: do we have a written incident response plan specifically covering synthetic identity fraud?
For most organizations, the honest answer to all three is no. That is the gap that needs to close — not next year, not after the next incident. The next deepfake call your finance team receives may be indistinguishable from a legitimate one. The question is whether your architecture requires them to tell the difference.