A finance employee at the Hong Kong office of Arup, the British engineering firm behind the Sydney Opera House, logged into a video call in January 2024 expecting a routine conversation about a confidential transaction. He saw his CFO. He saw colleagues he recognized. Everyone on the call looked and sounded exactly right. By the end of that meeting, he'd authorized 15 wire transfers totaling roughly $25.6 million to accounts in Hong Kong. Nobody on that call was real. Every face and every voice had been built from publicly available video and audio of Arup executives, harvested from webinars, press interviews, and old conference recordings. Deepfake fraud business risk isn't some future threat anymore. It already cost one company more than $25 million, and the case is still being cited as a wake-up call two years later.
Here's the part that should worry every CFO reading this: nothing about Arup's internal systems was breached. No malware, no stolen credentials, no network intrusion. The company's own chief information officer, Rob Greig, later described it as social engineering enhanced by technology, not a hack in any traditional sense. The scam walked in through a video call, wearing faces everyone trusted.
The attack didn't begin with a dramatic deepfake. It started small, the way most of these things do: a phishing email that appeared to come from the company's UK-based CFO, asking the employee to handle a private transaction. The employee was suspicious enough to ask for a video call to confirm the request before acting on it. Good instinct. Standard security training tells people to verify.
That's exactly where it fell apart. On the call, the "CFO" looked right. He sounded right. Several other familiar colleagues appeared to corroborate the request in real time. Multiple synchronized deepfakes, all speaking and gesturing naturally, did something a single spoofed email never could: they overwhelmed the employee's skepticism with what felt like overwhelming, in-person proof. He stopped questioning and started executing transfers.
CNN's reporting on the case noted that Arup's spokesperson confirmed fake voices and images were used, and that the company's broader financial stability wasn't affected, small comfort for whoever had to explain the loss internally. The fraud only surfaced when the employee later checked in with Arup's actual headquarters about the "secret" transaction and was told, flatly, that no such meeting had ever happened. By then the money was gone. As of this writing, no suspects have been publicly identified and the funds haven't been recovered.
It's tempting to file Arup under "extreme case, happened overseas, unlikely to hit a mid-size US company." That would be a mistake. The tools that made the Arup attack possible have gotten cheaper and more accessible since 2024, not less. Real-time video deepfake generation, voice cloning from a few seconds of audio, AI-written phishing emails with none of the broken grammar that used to give scams away: all of it is now available at a fraction of the cost and technical skill it required even a couple of years ago.
Deloitte's Center for Financial Services has been tracking this trend closely, and their projections are blunt. The firm estimates that generative AI-enabled fraud losses in the United States could climb from about $12.3 billion in 2023 to as much as $40 billion by 2027, a compound annual growth rate north of 30 percent. Deloitte's analysts specifically flagged attacks like the Arup incident as a pattern likely to proliferate as bad actors gain access to increasingly sophisticated, yet affordable, generative AI tools. That's not a niche warning aimed at big banks. Small and midsize businesses, particularly ones without dedicated fraud teams, are arguably softer targets.
Business email compromise, the less flashy cousin of deepfake fraud, already shows what happens when scammers scale a working formula. The FBI's Internet Crime Complaint Center logged more than 21,000 business email compromise complaints in 2024 alone, with reported losses near $2.8 billion, part of a broader total of $16.6 billion in cybercrime losses reported to the FBI's IC3 that year. Add generative AI to that formula, letting scammers write flawless, context-aware emails and then back them up with a cloned voice or a live deepfake video call, and you get an attack that's both more convincing and easier to run at scale.
Full real-time video deepfakes still take some setup. Voice cloning doesn't. Researchers have found that a few seconds of someone's recorded speech, the kind of clip anyone might have from a company podcast, a conference talk, or a LinkedIn video, is enough to generate a convincing voice clone. That's a low bar for a scammer targeting a specific executive.
The typical playbook looks familiar even without AI involved: an urgent call, supposedly from the CEO or a vendor, asking someone in finance to push through a wire transfer, change banking details, or share a one-time passcode before anyone has time to think it through. What's changed is that the voice on the other end can now sound exactly like the person it's impersonating, not a rough approximation. The Federal Trade Commission has been sounding the alarm on this specifically. Its Voice Cloning Challenge and related consumer guidance point directly at scenarios where scammers clone the voice of a CEO or other executive to trick employees into transferring funds or paying a fake invoice. That's not a hypothetical use case anymore. It's the exact mechanism regulators are already trying to build defenses against.
Here's the uncomfortable truth about CEO fraud prevention in the deepfake era: awareness alone won't cut it. Telling employees to "watch for red flags" made sense when deepfakes had glitchy edges and stilted voices. Those tells are disappearing fast. Arup's own CIO reportedly tried deep faking himself using free, open-source tools after the incident and had something passable within 45 minutes. If a company's chief information officer can do that on a lunch break, betting your fraud defense on an employee's ability to spot a fake in real time isn't a strategy anymore.
What works instead is process, not vigilance. A short list, because this genuinely is the kind of thing that benefits from a checklist rather than a paragraph:
Require a callback to a known, independently verified phone number, never one provided during the suspicious call or email itself, before executing any large transfer.
Build in a mandatory delay period, even 24 hours, for transfers above a set threshold, so there's time to catch inconsistencies.
Separate the person requesting a transfer from the person with authority to approve it, so no single employee (or single deepfake call) can push money out the door alone.
None of that requires exotic technology. It requires a policy that survives contact with a convincing fake, and enough internal discipline to follow it even when the person on the video call looks exactly like the boss.
For companies without an in-house security team big enough to build and enforce these controls, this is often where outside help earns its keep. InfineneTech.com works with midsize businesses to put verification workflows, secure communication channels, and staff training programs in place before a deepfake call ever reaches someone's inbox, rather than reacting after a wire transfer has already gone out. It's less about buying a tool and more about closing the process gaps that made the Arup case possible in the first place.
Most cybersecurity training programs still lean heavily on spotting suspicious emails: bad grammar, mismatched sender addresses, urgent language. Those lessons aren't wrong, they're just incomplete now. Cybersecurity training 2026 needs to assume that the phishing email is only step one of a multi-channel attack, one that might escalate to a phone call with a cloned voice or a video meeting with a synthetic face.
Effective training in this environment does a few things differently. It runs live simulations, not just slide decks, so employees experience the pressure of an "urgent" request from someone who sounds exactly right. It teaches specific verification habits (the callback, the second approver) as muscle memory, not optional advice. And it extends past the finance team, since deepfake fraud increasingly targets HR, IT help desks, and even customer support staff who might be tricked into resetting credentials for a "locked-out executive." Training that only covers wire fraud misses a growing share of the attack surface.
There's a financial wrinkle here that catches a lot of companies off guard: not every cyber insurance policy automatically covers social engineering losses the way it covers a ransomware attack or a data breach. Social engineering fraud, including deepfake-enabled wire transfers, is frequently written as a separate rider with its own (often much lower) coverage limit. A business that assumes its general cyber policy has it covered may find out otherwise only after the money is already gone.
It's worth a direct conversation with a broker about exactly what's covered, what the sub-limits are for social engineering and funds-transfer fraud specifically, and what documentation an insurer will require to pay a claim. Given that the FBI's own Recovery Asset Team reported freezing a meaningful share of fraudulent BEC transfers in 2024 when notified quickly, speed of reporting matters almost as much as the policy language itself.
Sometimes, but the window is closing fast. A year or two ago, security teams taught employees to watch for stiff facial movement, mismatched lighting, or a voice that lagged slightly behind lip movement. Those tells still show up occasionally, especially in cheaper, less-resourced attacks. Ask someone to turn their head in profile, and some deepfake tools still struggle to render the side of a face convincingly. Ask them to pick up a pencil or change the light source, and a poorly built synthetic feed can glitch or blur.
None of that is a reliable long-term defense, though. The Arup case is instructive here precisely because the deepfakes were good enough to fool someone who'd already been suspicious enough to request a video call as verification. That's the whole problem with treating human detection as a primary control: it only works until the technology improves past whatever tell you were trained to spot, and right now the technology is improving quickly. A company that's betting its entire defense on employees noticing a slightly off blink rate is going to lose that bet within a year or two, if it hasn't already.
The more durable approach treats every unusual, urgent, high-stakes request the same way, regardless of how convincing the person asking appears to be. If a request involves money leaving the company or sensitive data going out the door, it gets verified through a separate, pre-established channel. Full stop. That's not a technology fix. It's a cultural one, and it's the kind of policy that keeps working even as the deepfakes themselves keep getting better.
There's a temptation to treat this as purely an IT problem, something to hand off to the security team and forget about. That's a mistake, because the actual point of failure in almost every deepfake fraud case so far hasn't been a technology gap. It's been a moment where a single employee, under pressure and facing what looked like undeniable proof, made a judgment call alone.
Fixing that means giving employees explicit permission to slow down, even when the person on the other end is, apparently, the CEO and clearly in a hurry. Companies that build a genuine no-blame culture around pausing a suspicious transfer, rather than quietly punishing the employee who asks one too many questions, tend to catch these attempts before money moves. The finance employee at Arup wasn't reckless. He followed a reasonable instinct to verify by video call. The lesson isn't that he should have been more suspicious. It's that the verification method itself needs to sit outside whatever channel the request arrived through in the first place.
Deepfake fraud uses AI-generated video, audio, or images to impersonate a real person, usually an executive, vendor, or colleague, convincingly enough to trick someone into transferring money, sharing credentials, or approving a fraudulent request. It ranges from a cloned voice on a phone call to a fully synthetic video participant on a conference call, as seen in the Arup case.
Most attacks start with reconnaissance: scammers pull publicly available video, audio, and organizational details from a company's website, LinkedIn profiles, webinars, or press appearances. That material trains the deepfake, and the org chart tells them exactly who to impersonate and who to target with the resulting fake.
Put a callback verification policy in place for any large fund transfer or banking-detail change request, using a phone number the company already has on file, never a number provided in the suspicious message itself. It's low-cost, fast to implement, and it would have stopped the Arup fraud outright.
Sometimes, but not automatically. Social engineering and funds-transfer fraud are often covered under a separate rider with a lower payout limit than the main cyber policy, so businesses should confirm the specifics with their broker rather than assume broad coverage applies.
Contact the receiving bank immediately to request a transfer recall, then file a report with the FBI's Internet Crime Complaint Center at ic3.gov, since fast reporting significantly improves the odds of freezing funds before they move again. Loop in legal counsel and, if one exists, the company's cyber insurance carrier the same day.
Deepfake fraud isn't a distant, futuristic threat anymore. It's a line item in the FBI's crime statistics and a $25 million lesson one engineering firm would rather not have learned. The businesses that come out ahead of this won't be the ones with the flashiest detection software. They'll be the ones that built boring, unglamorous verification habits into their process before they needed them.
