Scammers used AI voices to impersonate CEOs in $129M fraud
How did Soall, Hazan, and Sheetrit pull off a $129M AI impersonation scheme, and why are so many companies still falling for it?
A radio-style debate on a real news story, written for C2 learners of English (proficient): native-level debate: fast, idiomatic and with no simplification.
Hear the whole debate in the app Free every day: today's episode at your level, with a synced transcript and the key vocabulary.
Get it on Google Play Download on the App StoreTeacher? Printable English C2 worksheet for this episode (questions, gap fill, answer key).
The debate in brief
Three men used AI-cloned CEO voices paired with impersonated lawyers to deceive thousands of companies into transferring $129 million. The debate centers on whether defenses should rely on procedural rules like callbacks, staff training to resist social engineering, or structural barriers that remove human decision-making from the process.
Key vocabulary
| English | Meaning and example from the episode |
|---|---|
| impersonate | pretend to be someone else to deceive them“Three men allegedly stole $129 million using AI voices to impersonate CEOs.” |
| fintech | financial technology company or sector“Hospitals in Missouri, a fintech in London.” |
| undetected | not discovered or noticed by anyone“July 2024 to February 2026, undetected.” |
| callback | return call to verify identity by ringing back“Callback on a known number, every time.” |
| deepfake | synthetic media made with AI to imitate someone“Sarah, does Marcus's gambit here—that 'confidential' kills the callback reflex—square with what you've seen in deepfake forensics?” |
Who's talking
- Host
- Marcus Webb
Former corporate security director
Authority stacking—particularly the fake lawyer blessing the wire—made the scheme work. Rules break when invoked with confidentiality claims, so structural barriers like mandatory dual signatories are needed over procedural ones. - Sarah Chen
AI ethics researcher and policy advocate
The cloner was bought off-the-shelf; the real vulnerability is that voice alone needs ninety seconds of plausibility to work. Callbacks must be unconditional rules, not suspicion-based, but adding more people just scales the impersonation problem. - David Okonkwo
In-house compliance officer at a regional
Policies existed on paper but failed in execution because staff never drilled them under pressure. Callbacks work as plumbing—a practiced step—regardless of charm or urgency; the fix is routine drills and checking at every transfer stage.
The panelists are fictional characters, and their voices are generated with AI.
What they debate
- Did Soall, Hazan, and Sheetrit build the AI tools themselves or buy them off-the-shelf?
- Should banks legally block 'confidential deal' wire requests before they happen?
- Why did finance staff at hospitals and banks trust AI voices over their own verification instincts?
Transcript (beginning)
Welcome to RoundTalk. Three men allegedly stole $129 million using AI voices to impersonate CEOs. They're caught, but thousands of companies fell for it. How did they pull it off, and why are our defenses so brittle? Marcus, corporate security consultant; Sarah, AI ethics researcher; David, bank compliance officer.
Marcus, you've spent eighteen years watching finance teams under pressure. What made Soall, Hazan, and Sheetrit's calls so effective?
The tools are beside the point. What made those calls work was the second voice—the fake British lawyer at a big firm, blessing the wire. Authority stacked on authority. That's where it breaks down.
Sarah, Marcus says the toolkit's almost irrelevant—it's the social engineering that scales. Did Soall and his partners build their own voice cloner, or just licence something off-the-shelf?
Bought, almost certainly. Nobody in that indictment is described as an engineer. And that's the problem nobody wants to talk about—somebody built that cloner, somebody sold it, and they're not in the dock in Manhattan. Soall rented a voice. July 2024 to February 2026, undetected.
David, you've trained hundreds of staff on social engineering. When a voice sounds that convincing—the CEO, then the lawyer—do your teams even have a fighting chance?
A fighting chance, yes—but not by ear. Last year a clerk of mine got a "CFO" call, perfect voice. She hung up and rang the desk extension. That's it. Callback on a known number, every time.
No, wait—callback assumes she thinks something's wrong.
Right, Sarah.
That's exactly why it can't depend on suspicion. Make the callback unconditional—every wire, every time, no judgement call. David's clerk shouldn't need a hunch. Soall's whole business model was manufacturing urgency so nobody thinks anything's wrong.
Wrong premise, Marcus. Callback isn't a suspicion tool, it's a plumbing tool. My clerk didn't smell a rat—she followed the step. Make it a rule and you don't need instinct.
Continue listening: 26 more turns The full episode, with the transcript synced to the audio, is in the RoundTalk app.
Get it on Google Play Download on the App Store