
AI voice cloning lets a fraudster imitate the voice of a relative or a colleague from a short audio clip. In practice, recognizing a voice no longer proves anything. The two defenses that hold are a verbal password agreed in advance and a callback through a channel you control.
This article explains what AI changes about vishing, what it does not change, and the habits to adopt. It complements our guide to security awareness training and vishing.
An AI model learns the characteristics of a voice: timbre, pace, intonation. It can then make that voice say any text. Sometimes a few seconds of recording are enough, taken from a video posted online, a voice message or a webinar.
The US Federal Trade Commission has been warning about this since 2023. According to the FTC, scammers use voice cloning to make their requests for money or information more believable. It gives the example of a relative who seems to be calling in an emergency, with a voice that sounds like theirs.
For an organization, the exposure is real. The voices of executives, spokespeople and salespeople circulate in videos, podcasts and conferences. That does not mean every suspicious call is a clone. It means the possibility must stay open, especially when the request is unusual, urgent or confidential.
Vishing relies on trust and pressure. AI does not change that pattern. It removes one of the clues employees used to rely on: “that is not his voice”.
Here is the good news: the content of the request remains the best clue. Whatever the voice, a call that demands, under pressure, an off-system wire transfer, a code or access is handled the same way.
The most cited case involves the engineering firm Arup in Hong Kong. In 2024, an employee took part in a video conference where he believed he saw the chief financial officer and colleagues. All of them were AI recreations, according to Hong Kong police as reported by CNN.
The employee made 15 wire transfers for a total of 200 million Hong Kong dollars, around 25.6 million US dollars. The fraud was only discovered when he checked with the company's head office.
This case is a video conference, not a voice call. It still shows the same mechanism: resemblance outweighed procedure. Verification through an independent channel revealed the fraud, but too late.
Three reasons explain why the ear is no longer a reliable test.
Asking employees to “spot the clone” means giving them a task nobody performs reliably. It is better to replace detection with a procedure.
This is a word or phrase agreed in advance between two people, never written in a message, never used anywhere else. It suits small circles: a management team and its treasury, an executive and their assistant, a family. It is renewed regularly.
“Before we go any further, give me our agreed word.”
Its limit is organizational: it does not scale to a whole company and is lost when someone forgets it. Keep it for high-stakes circuits.
This is the most general defense. You hang up, then call the person back on a number you already hold: internal directory, bank card, contract. The method is detailed in our article on the callback procedure through an independent channel.
The FTC gives the same advice to individuals: do not trust the voice, call the person on a number you know is safe, and go through other relatives if they do not answer.
Here is a simple course of action, valid with or without cloning.
For finance teams, the control must be structural: dual approval, thresholds, delays. A voice clone does not get around a procedure applied to everyone. That is also what our article on CEO fraud recommends.
If the voice proves nothing, the context of the call still speaks. A clone is a tool, not a complete scenario. The attacker must still obtain an action, and that action gives them away.
These signals are learned by speaking. The more an employee has heard them in conditions close to the real thing, the faster they spot them when they arrive.
An instruction that depends on individual vigilance wears out quickly. The most solid defenses are built into how the company operates.
For IT support, a particular case is covered in our article on social engineering at the help desk.
Knowing that a voice can be fake is not enough to act. Faced with a well-known boss who is getting annoyed, politeness takes over. The reflex of hanging up and calling back cannot be made up on the spot.
A role-play lets people live this situation before it happens. With Face Up, the trainer writes a scenario in which a persona played by an AI poses as a hurried superior. The learner speaks out loud, applies the procedure or not, then receives a debrief with the sentences they actually said. They can replay until the wording comes without effort.
Not reliably by ear. Hesitations or odd intonation may exist, but their absence proves nothing. It is better to rely on a verification procedure, the same for every sensitive call, whatever the voice.
It is useful for a small circle, provided it is secret and renewed. It does not replace the callback or dual approval for sensitive transactions.
The tools are available everywhere. In its 2025 report, Cybermalveillance.gouv.fr, the French national cybersecurity support service, notes a 517% rise in phone number spoofing, which often accompanies these calls. It is prudent to prepare teams now.
To train your teams to hold the callback against a familiar voice, discover Face Up role-plays or book a demo.
From reading to practice.
Pick a real-life situation and talk to an AI persona, by video with an avatar or by audio call. 5 to 10 minutes, no account, instant debrief.
Try a scenario →