At a glance
- Who should read this
- Finance, treasury and accounts payable leaders, executive assistants, fraud teams and security awareness leads.
- The risk
- A short sample of someone's voice can now produce a convincing clone. Any process that treats a familiar voice, or a video call, as proof of identity for a payment or account change can be exploited.
- Priority actions
- Verify every high-risk request by calling back on contact details you already hold, never those supplied with the request.
- Require independent second approval for payments and bank-detail changes.
- Give staff written permission to pause any request, however senior it appears.
Summary
Voice cloning has moved impersonation fraud from email into phone calls, voice notes and video meetings. Business email compromise and executive impersonation are long-established, high-loss fraud types, and synthetic voice and video add a layer of apparent proof that many approval processes were never designed to question.
In a widely reported case in early 2024, an employee in the Hong Kong office of the engineering firm Arup transferred about US$25 million after a video call in which, according to police, the other participants were deepfakes of the firm's chief financial officer and colleagues. US authorities, including the NSA, FBI, CISA and FinCEN, have warned about synthetic voice and video being used in fraud and impersonation.
The defense does not depend on detecting the fake. It depends on processes in which a convincing voice is not enough.
Attack patterns
| Pattern | How it works | What it exploits |
|---|---|---|
| Urgent executive request | A call or voice note in a senior leader's voice asks for a confidential transfer, often before a deadline | Authority, urgency, secrecy |
| Staged video meeting | A video call with one or more synthetic participants is used to remove doubt raised by an earlier message | Trust in seeing familiar faces |
| Supplier bank-detail change | A call from an apparent supplier contact confirms new bank details sent by email, sometimes from a genuinely compromised account | Verification that uses details supplied by the requester |
| Approval by phone | A caller impersonates an approver to authorize a payment that is already in the system | Processes that accept verbal approval |
| Help desk reset | A caller in an employee's voice asks the service desk to reset multi-factor authentication or a password | Voice and personal facts used as identity checks |
Fraudsters commonly combine channels: an email or message sets up the request, and the voice or video contact is used to overcome hesitation. They research their targets in advance, using public recordings, social media and earlier compromises to make requests specific and plausible.
Why detection is not the answer
Listening for artifacts, such as flat intonation or odd pauses, sometimes catches a poor clone, and context signals such as a new number, pressure for secrecy or resistance to a callback are valuable warnings. But clone quality is improving quickly, phone and conferencing compression hides artifacts, and the most dangerous attempts are the ones that sound right.
Treat recognition as a tripwire: if something seems wrong, stop. If it seems right, verify anyway. No impression of a voice, and no detection tool score, should authorize a payment.
Controls that break the attack
Out-of-band verification
Before acting on a high-risk request, the recipient starts a new contact with the requester through a channel and contact details that existed before the request, such as the company directory or a vendor record verified at onboarding. Numbers, links or contacts supplied in the request are never used. An email saying "the CFO will call you to confirm" is not verification, because the fraudster chose the channel.
Rule-based triggers
Verification is triggered by the request, not by suspicion. Typical triggers include any payment to a new payee, payments above a threshold or outside normal timing, any change to bank details, and any request for urgency, secrecy or an exception to process.
Independent second approval
A second approver verifies independently rather than relying on the same email thread or caller. Segregation of duties ensures that the person who creates or changes a payee cannot also approve or release payments to it.
Holds and notifications
Changes to bank details take effect only after a hold period, and a notice is sent to the existing contact through a separate channel, so a genuine supplier can object. The first payments to a new account are monitored.
Permission to pause
Leadership states in writing that no one will be criticized for verifying a request, even if it delays a genuine one, and executives agree in advance to be verified. A real executive will usually wait ten minutes. A fraudster will push harder, and that difference is itself a signal.
A note on code words
Shared code words can add a useful layer, especially for small teams and families, but they leak, are rarely changed and are easily forgotten under pressure. The verifier should ask for them, never accept them volunteered, and a wrong or missing answer should stop the request until a callback verifies it. A code word supplements out-of-band verification; it does not replace it.
If a payment has already been sent
- Contact your bank immediately to request a recall or freeze. Recovery chances fall quickly with time.
- Report to law enforcement. In the United States, file with the FBI's Internet Crime Complaint Center (IC3), whose Recovery Asset Team works with US banks to freeze funds sent to domestic accounts in some cases. Elsewhere, use your national equivalent.
- Preserve evidence: call logs, voice notes, messages and email headers.
- Warn other likely targets in the same team, since fraudsters often try several people.
- Review and close the process gap the fraud used.
References
- FinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004), FinCEN, 2024
- Contextualizing Deepfake Threats to Organizations, NSA, FBI and CISA, 2023
- UK multinational Arup confirmed as victim of HK$200 million deepfake scam, South China Morning Post, May 2024
- Internet Crime Complaint Center (IC3), FBI
This advisory draws on public research, standards and official guidance. It is general guidance, not advice for a specific environment. Related training: Synthetic Media and Fraud Defense.