1 in 5 CFOs Fear Deepfake Impersonation. Half Can't Catch It.

In February 2024, an employee at the engineering firm Arup wired $25 million after joining a video call where every other participant, including a convincing replica of the CFO, was a deepfake. The employee saw a face, heard a voice, and followed instructions. Detection failed completely because the attack was designed to defeat it.

That case is no longer an outlier. In October 2026, Early Warning Services' fraud prevention unit Certos published findings from a survey of 230 senior financial leaders: 81% said their organization experienced attempted fraud involving AI-generated or AI-enhanced content in the past year. Eighty-four percent said AI-enabled fraud is harder to detect than traditional fraud. One in five named executive deepfake impersonation as their single biggest concern. And only 50% expressed confidence they could catch an AI fraud attempt before funds move.

The other half cannot.

Why Detection Is a Race Most Organizations Are Losing

Voice cloning now requires as little as 20 to 30 seconds of audio, and executives post earnings calls, conference keynotes, and podcast appearances publicly. Human detection accuracy for AI-generated audio hovers around 55 to 60%, barely above chance. Detection thresholds that worked in 2023 are not the thresholds needed in 2026.

The Certos data describes the attack pattern precisely. An impersonator poses as a senior executive, contacts a finance team member by phone or video call, and invokes urgency: a deal closing, a regulatory deadline, a confidential acquisition. The finance team member can comply quickly or slow the process down and risk looking obstructionist to someone who sounds exactly like the CFO. Most authorization processes are built around identity signals AI can now replicate: voice, caller ID, familiarity with how the executive speaks. These are no longer reliable controls.

The Operational Fix

The answer is not another layer of software trying to detect what an AI-generated voice sounds like. It is a shared secret the impersonator cannot access because it was never transmitted over any channel an attacker can intercept.

Trust Onion closes this gap with rotating codewords. Three words, generated locally on each device, change every 60 seconds, with no server and no internet required. When a caller claims to be the CFO authorizing a wire transfer, the finance team member asks one question: "What are the words?"

The real executive knows the current three words. An impersonator does not. The words were never spoken on a public call, never sent over email, and never stored on a server that could be breached. They expire automatically. Finance teams need to establish, before any attack occurs, that every executive request for a wire transfer, vendor banking change, or urgent payment will include this challenge, with no exceptions.

This is an operational policy, not a technology project. The CFOs who are confident they would catch an AI fraud attempt likely have some version of an out-of-band verification step already in place. The 50% who are not confident probably do not.

The Certos survey covers 230 senior financial leaders across a range of organization sizes. Eighty-one percent experienced attempted AI fraud in the past year, which means attempted fraud is now the norm, not the exception. Wire transfers above certain thresholds are often unrecoverable within hours. Vendor payment redirections can sit undetected for weeks.

The finance team member who receives the call is not naive. They are operating without the tools to verify the caller's identity. That is an organizational gap, not a personal one.

AI can clone a voice. It cannot produce three words that change every 60 seconds, were never transmitted externally, and exist only on the devices of verified personnel.

A controlled voice-cloning safety demonstration

Try our tool to see just how easy voice cloning is. This controlled safety demonstration exists only to show how easy voice cloning is and to make the risk tangible. It is not a real incident or proof of anyone's identity. Your submitted sample is handled transiently to generate the fixed demonstration audio. Trust Onion does not identify you or the person from the sample, and does not keep the submitted recording or generated audio. The temporary voice clone is deleted from our voice provider as soon as the demo finishes; if that deletion fails, it is retried automatically until the provider confirms it.