The fake executive email or phone call demanding an urgent transfer is an old fraud method. Generative AI has added two tools to it: voice cloning, which can say any text in a voice learned from short recordings, and the video deepfake, which maps one person's face and facial expressions onto another person's movements. In December 2024 the FBI's Internet Crime Complaint Center (IC3) described in a public warning how fraudsters use voice and video created with generative AI to impersonate relatives, company executives and officials.
What voice and video forgery can do today
A voice cloning model learns a person's timbre, intonation and speaking pace from recordings of them. It can be used in two ways. In text-to-speech mode the fraudster types a sentence and the model reads it out in the target's voice; this is how a voice message or a pre-scripted call is made. With real-time voice conversion the fraudster does the talking and the system converts their voice with a short delay, so a live, back-and-forth conversation is possible.
Getting a sample does not require hacking anything. There is usually public footage of an executive: interviews, conference talks, webinars, company videos, speeches shared on social media. A voicemail greeting or an earlier recorded phone call can also be enough.
The live form of video forgery is webcam face swapping: in the call, the target's face appears in place of the fraudster's. The image can distort during quick head turns, in profile, when a hand or an object passes in front of the face, and in poor lighting. Video call compression and low resolution, however, blur exactly these flaws. The written part is no obstacle either: a language model writes a flawless email in any language, imitating the style of the impersonated executive.
| Format | Typical use | Where it can slip |
|---|---|---|
| Pre-generated voice message | Voicemail, messaging app, short call | Cannot respond to follow-up questions; the intonation stays even in a tense situation |
| Real-time voice conversion | Live phone call | Latency, odd pauses, voice distortion |
| Live face swap in a video call | “Executive” meeting on camera | Profile view, a hand in front of the face, fast movement, changes in lighting |
| Generated written text | Preparatory or confirming email | Only its content can give it away |
The typical scenario: an urgent instruction from an executive
The fake executive call is a form of social engineering, and its steps are quite predictable.
- Reconnaissance. From public sources the fraudster works out who the finance lead is, who can initiate transfers, who the main suppliers are, and when the person they plan to impersonate is traveling or on leave. That person is harder to reach directly at such times, which works in the fraudster's favor.
- Preparation. An email arrives from a lookalike domain or a compromised mailbox announcing the call: a confidential matter, someone will be in touch shortly.
- The call. The “executive” phones, or sets up a video call where there may even be several fake participants. The story is a confidential acquisition, a regulatory proceeding, an overdue supplier payment or a contract that has to be closed that same day.
- Pressure. A tight deadline, a request for secrecy (“don't tell anyone about this”), the weight of authority, and a way around the usual procedure (“I don't have time to approve it in the system right now, just send the transfer and we'll sort out the approval afterwards”).
- The transfer. The money goes to a new, previously unused account, often in several installments, and the fraudsters quickly move it on from there.
The same pattern works for other goals too. A fake IT staff member asks the help desk for a password reset or an MFA code, a fake supplier announces a new bank account number, and a fake relative, posing as a family member in trouble, asks a private individual for money.
Why your ears and eyes are not enough
Most people believe they would recognize a familiar voice. During a call, though, attention goes to what is being said and hardly at all to the technical features of the voice. If the voice is familiar and the story is plausible, the brain fills in the missing details, and stress and haste make this worse. On top of that, the narrow bandwidth of a phone line and the compression used in calls remove precisely the subtle differences that would give away a synthetic voice.
Technical signals are no more reliable. Caller ID can be spoofed, so the display may even show the executive's real number. The email may come from the executive's genuine but compromised mailbox. Deepfake detection software gives only a probability, in live calls too, and has to be retrained continually as generation models improve. Its output is one signal among many, and a decision cannot rest on it.
There are telltale signs, and they are worth knowing:
- the caller avoids being called back, or insists on a new phone number;
- the camera does not come on, or only briefly, or the picture is poor, blamed on “a bad connection”;
- an unexpected question is followed by a noticeably long pause, or the answer does not fit the question;
- the background is suspiciously quiet or, on the contrary, too noisy;
- the request differs from how that executive normally handles such matters.
Their absence, however, proves nothing. The defense therefore has to sit where the authenticity of the voice and the image does not matter: in the business process itself.
Process controls: callback, challenge question, dual approval
A process control protects against deepfakes if the fraudster cannot get past it even with a perfect voice and face. The proven elements are simple; the hard part is applying them without exception.
Callback on a known number
Every request that involves moving money, changing a bank account number, resetting a password or granting access has to be confirmed over a separate, independent channel. In practice the person who receives the request hangs up and calls the requester back on the number recorded in the internal phone directory or the supplier master data. A number given in the call, the email or the chat message cannot be used for this, because the fraudster supplied it.
Challenge question and code word
Executives with authority over financial decisions and the finance staff can agree in advance on a code word or a challenge question to use when a request is unusual. The answer should not be something that can be worked out from public data, social media or the org chart, it should never be written in an email or a chat, and it has to be changed when someone leaves the group. The IC3 warning mentioned above recommends the same to families: a secret word that only the family knows.
Dual approval
Above a set amount, two people approve the transfer, and the banking or finance system enforces this technically. The person who enters the request cannot be an approver. Adding a new beneficiary or changing an existing beneficiary's bank account number should be a separate process, with its own approval and a verification callback, before the first transfer goes out.
A verbal instruction alone is not enough
Make it a written rule that a phone or video call alone cannot trigger a transfer or change master data. The same applies at the IT help desk: nobody gets a password reset or a new MFA device registered on the strength of their voice, only after the established identification procedure. How well each MFA method holds up against tricks like these is covered in the article on MFA types.
What awareness training should teach about this
The goal of the training is for everyone to know by heart in which situations to hang up and call back, whom to call, on which number, and whom to tell afterwards. A short demonstration that plays a cloned voice, made with the consent of an executive taking part, shows how convincing the technology is. Participants then practice the callback in a role play: one asks for an urgent transfer, the other hangs up and checks the request on the recorded number.
The target groups:
- finance and procurement: transfers, new beneficiaries, bank account changes;
- HR and payroll: an “urgent” change to an employee's bank account number;
- IT help desk: requests for a password reset or a new MFA device over the phone;
- executive assistants: they often handle matters on the executive's behalf, so fraudsters refer to them;
- the executives themselves: the samples come from their public recordings, and their example decides whether the rule is taken seriously.
Voice fraud exercises (vishing simulations) are best run with rules agreed in advance and with management approval. The results should show the gaps in the process without shaming individual employees. Reporting a suspicious call also has to be easy: a known phone number or mailbox that anyone can contact at any time. Suspicious emails can be reported there as well; phishing by email has its own article.
If a call seems suspicious
- Do not transfer money, and do not give out a code, password or account number during the call.
- End the conversation politely, for example by saying you will call back. A real executive will understand.
- Call the requester back on the number recorded in the internal phone directory or the master data, and ask about the request.
- Report the incident to the person responsible for security or IT and to the head of finance, even if no damage was done. Others may have received a similar call.
- Keep the evidence: the call log with the number and time, related emails with full headers, the voice message or the chat.
- If the money has already gone, call the bank immediately and ask for a recall of the transfer. Minutes count here, because fraudsters move the funds on quickly. Then file a report with the police.
- If credentials were given out during the call, change the password immediately, revoke active sessions, and check the MFA devices registered to the account.
The impersonated executive has to be told as well, because their voice or name may turn up in a later attempt, and they know best which of their public recordings could have served as a sample.
Related service: Security and AI usage training →