From Xenovia LabsIssue 3, October 2026
Secure Enough

Probably fine.

ReportBreaches

Voice cloning fraud has become a line item

Fraud teams now plan for synthetic voices the way they once planned for forged signatures, and the controls that work are older than the threat.

By Marguerite Okonkwo-Hale, Senior Breach Correspondent3 min read

Two years ago, a cloned voice on a phone call was a conference demo. This year it is a column in corporate fraud budgets, sitting alongside check fraud and account takeover, and in conversations with fraud investigators, bank operations staff, and corporate treasurers over the past month, nobody described it as novel anymore. They described it as Tuesday.

The mechanics are simple enough to explain in a paragraph. Modern speech-synthesis models can produce a convincing clone of a specific voice from a short sample, often under a minute of clean audio. Executives, salespeople, and anyone who has ever recorded a podcast, a webinar, or a voicemail greeting has provided that sample. Real-time systems can take typed or spoken input and render it in the target voice with latency low enough for a phone conversation, and increasingly for video calls with a synthesized face to match.

What has changed is not that this is possible but that it is cheap, packaged, and sold to criminals as a service alongside the rest of the fraud toolkit.

Where it is being used

Three patterns dominate.

Executive impersonation. The classic business email compromise, in which a finance employee receives an urgent payment request from a spoofed executive, now arrives by voice or video. The call lends the request an authority an email cannot, and it defeats the standard training advice to "call and verify" when the victim believes they are already on a call with the person.

Help desk and account recovery. Attackers call a company's IT help desk or a bank's customer line impersonating an employee or customer, sometimes using a cloned voice of the real person, to reset passwords or re-enroll multi-factor devices. Help desks are measured on resolution speed and are, by design, helpful.

Family emergency scams. A relative receives a call from a child or grandchild in distress, asking for bail money or a hospital payment. The voice is cloned from social media video. These are lower in value and vastly higher in volume.

Why it works

Investigators are consistent on this point: the deepfake is rarely what makes the fraud succeed. It removes the last reason to doubt a request that was already engineered to bypass judgment, using the same ingredients fraud has always used: authority, urgency, secrecy, and a plausible reason the normal process does not apply this time.

"People keep asking me how to detect the fake voice. The honest answer is that in the moment, on a bad phone line, under time pressure, you mostly cannot. So stop building your controls around detection." — Fraud operations lead at a regional bank, speaking on background

Detection tools exist, and some are useful for screening recorded calls after the fact. None should be the thing standing between a cloned voice and a wire transfer.

What is working

The controls that defeat voice fraud are, for the most part, the controls that defeated fax fraud.

  • Out-of-band verification. Any request to move money, change payment details, or reset credentials is confirmed through a separate channel the requester did not initiate: a call back to a number already on file, a message to a known address. Not a reply to the call. Not a number provided during it.
  • Dual control that is actually independent. Two approvers who each verify separately. Two approvers watching the same video call provide one approval.
  • Mandatory delays for new payees and payment changes. Fraud is a race; a 24-hour hold wins it more often than any detector.
  • Help desk procedures that do not rely on voice. Identity verification through a registered device, a manager callback, or in-person confirmation for high-risk resets, regardless of how familiar the caller sounds.
  • Shared secrets for families and small teams. A code word agreed in advance is low-tech and surprisingly resilient.
  • Fast recall. Banks can sometimes claw back wires if notified within hours. Every organization should know, before an incident, exactly who to call.

The uncomfortable conclusion from this reporting is that the organizations getting hurt are not the ones that failed to buy deepfake detection. They are the ones whose payment processes could be completed by a single convinced person in a single afternoon. That was always the vulnerability. The voice is just the newest way to exploit it.

About the writer

Marguerite Okonkwo-Hale

Senior Breach Correspondent

Marguerite has read more breach notification letters than any living person and has started to dream in the passive voice. She covers incidents, disclosures, and the gap between the two.