Deepfake Voice Fraud: The CEO Scam Grows Up

Your bookkeeper takes a call. It is the owner, travelling, on a bad line, needing a payment released today because a deal is closing. The voice is right. The turns of phrase are right. The impatience is right.

It is not the owner. Thirty seconds of audio scraped from a conference talk, a podcast appearance, or a voicemail greeting is enough to clone a voice convincingly, and the tools to do it are neither expensive nor difficult.

This is not a new crime. It is the same impersonation fraud that has been draining business bank accounts for a decade, with the one remaining weak point removed.

Key takeaways:

• Everything we trained people to spot (clumsy grammar, odd phrasing, generic greetings) is exactly what AI fixes. Training built on typos is now worse than useless.
• Voice cloning needs only a few seconds of audio, and plenty of business owners have that much online without realizing.
• The defense is procedural, not technical: verify on a number you already had, never one supplied in the message.
• Agree a family and a company code word now. It costs nothing and it survives any improvement in the fakes.

What actually changed

The FBI’s Internet Crime Complaint Center issued a public warning about criminals using generative AI to commit fraud at larger scale, noting that these tools “reduce the time and effort criminals must expend to deceive their targets” and can “correct for human errors that might otherwise serve as warning signs of fraud”.

Sit with that second phrase, because it is the whole story. For years the practical advice was: look for the tells. Bad spelling, stilted English, a greeting that does not sound like the person. Those tells were never really about security, they were about the attacker being a stranger writing in a hurry, often in a second language.

Generative AI removes all of it. The email is fluent. The voice has the right accent. The video call shows a face that moves correctly. An organization whose entire defense is “our staff are good at spotting phishing” has a defense built on artifacts that no longer reliably appear.

The four ways it is being used

IC3’s list is worth reading as a list of things to expect rather than a curiosity.

Text. Criminals use AI-generated text “to appear believable to a reader in furtherance of social engineering”. This is the ordinary phishing and invoice fraud you already know, minus the giveaways. It layers straight onto business email compromise, where the attacker has already been reading the mailbox and knows exactly how your people write.

Voice. Criminals “generate short audio clips containing a loved one’s voice to impersonate a close relative in a crisis situation, asking for immediate financial assistance”. The business version substitutes an executive for the relative and an urgent payment for the bail money.

Video. Criminals “generate videos for real time video chats with alleged company executives, law enforcement, or other authority figures”. A video call used to be the way you confirmed someone was real. That assumption is no longer safe on its own.

Identity documents. Criminals “generate fraudulent identification documents, such as fake driver’s licenses or credentials”. This matters for anyone who verifies identity from a scanned document: new client onboarding, remote hiring, account changes.

Why this hits smaller businesses harder

A large company has a payments process with three people in it. A twelve-person company has a bookkeeper who does what the owner asks, because that is the job and because the owner is right there.

That closeness is normally an advantage. Here it is the vulnerability, because the entire control is one person’s judgment about whether a familiar voice sounds like itself. Add genuine urgency, real authority, and a plausible reason, and refusing feels like insubordination.

The fix is not to make your staff more suspicious of you. It is to remove the judgment call entirely.

The control that still works

Out-of-band verification. It is unglamorous, it is decades old, and it is untouched by any of this, because it does not depend on detecting the fake.

IC3’s own advice is to verify identity “by hanging up the phone, researching the contact of the bank or organization purporting to call you, and call the phone number directly”. The business version, written as policy:

Any payment, any change to bank details, any urgent transfer gets confirmed by voice on a number already on file. Not a number in the email. Not a number the caller gives you. The one you already had.

Then make refusing safe. Say it out loud, in a meeting, so everyone hears it: nobody will ever be in trouble for verifying a request from me. If the person who signs the cheques does not say that explicitly, the policy will lose to social pressure every single time.

The second control costs nothing. IC3 suggests you “create a secret word or phrase with your family to verify their identity”. Do the same at work. A word that is not written down anywhere, known to the handful of people who can move money, asked for when something feels off. A cloned voice does not know it.

What to do this month

Four things, none of which requires buying anything.

Write the callback rule down and apply it to the owner too. A rule with an exception at the top is not a rule.

Agree the code word with whoever can release funds, and one for your family while you are at it.

Check your own exposure. How much of your voice is publicly available? A recorded webinar, a podcast, a video on your website. You probably do not want to remove them, but knowing they exist changes how seriously you take the callback rule.

Tell clients how you will and will not contact them. Wire details never change by email, and you will always confirm by phone. That protects them from someone impersonating you, which is a call you would much rather not receive.

Then update the training. Retire “look for bad spelling” and replace it with “verify the request, not the message”. Our Spot the Phish quiz is a free way to see where your team currently sits, and it is a reasonable ten minutes at a staff meeting.

The honest limit

There is no product that reliably tells you a voice is synthetic, and be skeptical of anyone selling you one. Detection is losing an arms race it was never positioned to win.

What does work is layered and boring: strong sign-in so the mailbox does not get compromised in the first place, since most of these attacks start with a stolen password, payment processes that do not depend on recognizing a voice, and separation between who approves and who pays. Firms that move client money have the most to lose here, which is why we build the callback discipline into how we work with financial services clients.

If you would like us to look at how a fraudulent payment request would actually travel through your business, and where it would get stopped, that is part of a free assessment. Or email hello@cohesivesecurity.com and describe your current process. Most owners find the gap themselves partway through explaining it.