Deepfake and Voice-Cloning Fraud Against Organizations
How deepfake and voice-cloning fraud against companies actually works, why training staff to spot fakes is the wrong control, and the verification that holds.

The pattern that worries fraud teams is not a fake video posted online. It is a scheduled meeting that a finance employee is invited to, where the chief financial officer and two familiar colleagues appear on camera, the audio is a little compressed in the way all video calls are, and the only person in the call who is genuinely there is the employee being asked to make the transfer.
Everything in that scenario except the faces is a fraud technique that has existed for a decade. The research on the target, the plausible pretext, the manufactured urgency, the instruction to keep it confidential because of an acquisition, the insistence on handling it in this channel and not by email. The synthetic media is a delivery improvement to a known attack, and that framing is what tells you where to put your controls.
What the attacker has to assemble
This kind of fraud takes preparation, which is both bad news and good news. Bad because it means the approach will be well-informed, good because every stage leaves something you can disrupt.
- A target with authority to act. Someone who can release a payment, change a supplier record, reset a credential or create an account. Usually two or three levels below the executive being impersonated, which is deliberate: senior enough to act, junior enough not to question.
- An org chart. Who reports to whom, who approves what, who is currently away. Assembled from the company website, professional networks, press releases, conference programs and job adverts.
- Voice and likeness material. A short clean sample is enough for voice. Video comes from recorded talks, media appearances and the organization's own marketing. The NCSC's assessment of near-term AI effects is blunt about the direction of travel here.
- A pretext tied to something real. A genuine acquisition rumor, a real supplier, an actual audit, a known system migration. This is why reconnaissance matters more than the media quality.
- Channel control. The approach moves to where the attacker is strong and your process is weak: a personal mobile number, a messaging app, a meeting link they created, an out-of-hours call. The FBI's Internet Crime Complaint Center described exactly this progression in a campaign impersonating senior officials, where AI-generated voice messages were used to build rapport before the conversation moved to another platform.
Three failure points follow directly. Reduce what is publicly discoverable about approval chains. Make the channel itself a signal, so a request arriving outside the normal system is treated as unverified regardless of who appears to be making it. And remove the authority of a single person to complete a money movement alone.
Where it lands inside an organization
Payment fraud gets the headlines, and it is not the only destination.
Payment release and approval. A call that supplies the authority a payment process was waiting for, often with a story that explains why the usual approver is unreachable.
Supplier and payroll bank detail changes. Quieter and often larger, because the fraudulent account then receives legitimate invoices for months. A voice call is used to confirm an emailed change request, which is how the fraud defeats a control that assumed two channels meant two parties. Our article on invoice fraud and changed bank details covers the mechanics.
The help desk. An impersonated employee calls for a password reset or MFA re-enrollment. Voice recognition by a colleague has been doing identity proofing informally in many organizations for years, and it no longer can.
Recruitment and onboarding. A candidate who passes a video interview using a synthetic presentation, aiming at employment and the access that comes with it. If onboarding grants access on the strength of documents and a call, that is now a security process.
Extortion and reputation. Fabricated audio of an executive, used to pressure the organization directly or to move a share price. The joint NSA, FBI and CISA information sheet on deepfake threats treats brand and executive impersonation as organizational risks requiring preparation, not just awareness.
Why "train people to spot fakes" fails as a control
It is the instinctive response and it does not survive contact with how the fraud actually runs.
The artifacts move. Advice built on current tells, odd blinking, mismatched lip movement, flat intonation, a strange edge around the hairline, describes the output of particular tools at a particular moment. Teaching a checklist of artifacts creates confidence that expires silently.
Call quality hides everything. Compression, packet loss, a bad connection and a phone speaker remove exactly the detail a listener would need. An attacker does not need a perfect clone, only one that survives a mobile call, and "the line is terrible, can you hear me?" is a normal thing for a real executive to say.
Detection under pressure is not detection in a test. In a workshop, people told to look for fakes do moderately well. In a surprise call from someone three levels senior, during a deadline, where the implied cost of being wrong is accusing your boss of being an impostor, the same people do not perform the analysis at all. Europol's work on deepfakes and law enforcement makes the point that human detection is not a dependable basis for decisions.
It puts the burden in the wrong place. Asking an accounts payable clerk to make a forensic media judgment, alone, in real time, against a superior, is an unfair control. When it fails, and it will, the organization has also arranged for an individual to carry the blame for a process failure.
Liveness prompts are theater. Asking someone to turn their head, wave a hand or hold up a pen may catch a weak setup today. Any prompt that circulates widely stops working, and a confident attacker refuses, laughs it off, or blames bandwidth. Treat a passed check as no information.
The underlying error is building a control around recognition. Recognition is a human perceptual judgment, it degrades under stress, and the attacker gets to iterate until it fails. Verification is a process with a defined step, it produces the same answer tired or rushed, and it does not care how good the fake was.
Verification that survives a convincing impersonation
Callback to a number you looked up
The single highest-value rule. Any request that moves money, changes payment details or grants access is confirmed by calling the requester back on a number taken from your own directory or your own supplier record. Never a number in the message, the invite, the signature or the caller ID, and never by replying in the same channel. Write it as a rule with no exceptions rather than as guidance, because exceptions are what the pretext is designed to create.
Dual authorization with real separation
Payments above a threshold and every change to bank details require two people, and the second person must verify independently rather than approve on the first person's word. For a supplier change, that means contacting the supplier on previously held contact details, not the ones on the new invoice. The second approver should be able to refuse without it being awkward, which is a cultural requirement as much as a procedural one.
Identity proofing at the help desk that does not use voice
Resets should depend on something structural: a request through the ticketing system from a known account, verification through a manager who is contacted independently, a one-time code delivered to an enrolled device, or in-person confirmation. Remove "I recognized their voice" as an acceptable basis and say so in the procedure, so staff are not improvising.
A low-tech shared secret, used sparingly
For a small group with payment authority, a pre-agreed challenge phrase or a code word for genuine emergencies works precisely because it is not synthesizable. Keep it out of email and shared drives, agree how it is rotated, and understand it protects a narrow case rather than the whole process.
Channel rules stated in advance
Decide and publish that financial instructions never arrive by voice call, video call, or messaging app, full stop. When the channel itself is disqualifying, staff do not need to assess the content, and the executive whose likeness is being used is protected too.
Explicit permission to pause
Every one of these fraud attempts uses urgency and hierarchy together. A written, repeated, visibly executive-backed statement that nobody will be criticized for pausing a request to verify it is the control that makes the other controls usable. Without it, the procedure exists and gets skipped by the person who feels they cannot say no.
Making it real rather than written down
- Map the processes that move value or access and mark, for each one, what a determined caller would need to do to complete it. Where the answer is "convince one person", you have found your work.
- Write the callback and dual-authorization rules into the procedure itself, not into a training deck. The control should be a step in the system, ideally one that cannot be skipped without a record.
- Brief executives on their own role. They need to state that they expect to be called back and will never be annoyed by it. Reducing the volume of recorded voice and video they publish is also worth discussing, without pretending it can be eliminated.
- Test it with realistic exercises. Run a voice-based scenario against the payments process, with permission and a clear debrief, and measure whether the callback happened rather than whether anyone suspected a fake. Our guide to phishing simulation covers how to run these without damaging trust.
- Train the request, not the artifact. Staff should recognize the shape of a dangerous request: new payment details, urgency plus confidentiality, an unusual channel, pressure against the normal process. See training staff on payment fraud for the practical version.
- Agree the response in advance. Who is called first when a payment has gone out, which bank contact is used, what gets preserved, who speaks to the insurer. Minutes matter for recall.
Unsure about a message that came with the call? Paste the message or link into CyberWatch AI Scan for a free check.
If it has already happened
Contact the bank immediately and ask for a recall, before internal investigation, because the window for recovering funds is short. Preserve everything: the meeting invite and any recording, the calendar entry and who created it, call logs, the phone number or account used, the message thread, and the account details supplied. Do not delete the invite in embarrassment, as it is evidence and often the most attributable artifact in the whole episode. Report to law enforcement and, in the US, to the FBI's Internet Crime Complaint Center. Notify your insurer early, since many policies require prompt notification and may require their own responders. Then check whether the same approach was tried elsewhere in the organization, because it usually was. Our payment fraud response guide sets out the sequence, and the wider patterns sit in our guide to social engineering and our guide to AI-powered cyberattacks.
Frequently asked questions
How much audio does someone need to clone a voice?
Far less than most people assume. Current tools work from a short clip of clear speech, and that clip is usually already public in a conference recording, a webinar, a podcast appearance, a company video or a voicemail greeting. Treating an executive's recorded voice as confidential is no longer realistic.
Can staff be trained to spot a deepfake?
Not reliably, and it is the wrong thing to train. Quality improves continuously, artifacts vary by tool, and people perform much worse under time pressure from an apparent senior colleague than they do in a test. Train the verification process instead, which works whether or not the fake was good.
Do liveness checks like asking someone to turn their head still work?
They sometimes expose weaker setups and should never be your control. Any prompt that becomes widely known gets handled by the next generation of tools, and a confident attacker simply refuses or blames a bad connection. Treat a passed liveness check as no evidence at all.
What verification actually survives a convincing fake?
A callback on a number you look up yourself in your own directory, not one supplied in the request, combined with dual authorization for payments and any change to bank details. Both work because they do not depend on anyone judging whether a voice or face was real.
Our CEO travels and genuinely sends urgent requests. How do we handle that?
Agree the exception process in advance, in writing, with the executive's support. A standing rule that urgent out-of-channel requests are always confirmed by callback, and that the executive expects and welcomes it, removes the social pressure that makes the fraud work.
Which processes should we protect first?
Anything that moves money or grants access: payment release and approval, changes to supplier or payroll bank details, credential and MFA resets at the help desk, new user creation, and privileged access requests. These are where impersonation converts into loss.
Is this only a risk for large companies?
No. Smaller organizations are easier because the org chart is public, approval chains are short and informal, and one person often holds both the relationship and the payment authority. The material needed to impersonate a small company's owner is usually on its own website.
What should we do if we think a call was a deepfake?
Stop the transaction and contact the bank immediately if money moved, since fast recall is the only thing that recovers funds. Preserve the evidence, including the meeting invite, recording, call logs and the number or account used, then report it to law enforcement and your insurer.
Sources
- NSA, FBI, and CISA Release Cybersecurity Information Sheet on Deepfake Threats, CISA
- Facing reality? Law enforcement and the challenge of deepfakes, Europol
- Senior US Officials Impersonated in Malicious Messaging Campaign, FBI Internet Crime Complaint Center
- Home Page - Internet Crime Complaint Center (IC3), FBI Internet Crime Complaint Center
- The near-term impact of AI on the cyber threat, UK National Cyber Security Centre
- Cybersecurity for Small Business, US Federal Trade Commission
- Avoiding Social Engineering and Phishing Attacks, CISA
- Phishing scams: how to spot and report them, UK National Cyber Security Centre
- ChatGPT - the impact of Large Language Models on Law Enforcement, Europol


