On November 5, the US Attorney's Office for the Southern District of New York unsealed an indictment against three men. Prosecutors say they used AI to copy the voices and faces of company executives and law firm attorneys, and talked employees into wiring about $129 million to accounts the group controlled (DOJ press release).
Everything below is an allegation. The release says so itself: the charges are accusations, and the defendants are presumed innocent unless and until proven guilty. The Record reports that one defendant's lawyers deny his involvement (The Record). I'm using the case because the method is one a small company can actually defend against, and it has nothing to do with firewalls.
What prosecutors say happened
The three named are Jacob Soall, Sharon Hazan and Guy Sheetrit. Per the release, the alleged activity ran between 2024 and 2026. Soall was arrested in Spain on June 17, Hazan in Spain on August 5, and Sheetrit in the UK on September 7. Extradition requests are pending. Each faces seven counts, including wire fraud, aggravated identity theft and money laundering.
The method, as the indictment describes it:
- The group pretended to be the CEO of the target company and told an employee they needed help with a confidential financial transaction. On calls, they used AI to mimic the CEO's voice and, in some cases, image.
- The fake CEO sent the employee to lawyers at large international law firms, supposedly working on the same deal. Those lawyers were impersonated too.
- One of the fake lawyers told the employee to wire a large sum to a specific bank account. It belonged to the conspiracy.
- The money went to accounts in countries including the US, Spain, Portugal, Greece, the UAE, Singapore, Austria, Slovakia and China, and was then converted to cryptocurrency.
One detail matters for what you do about it. The release says Soall kept a handwritten script titled "Lawyer Call", and that it included language meant to convince employees not to discuss the conversation with colleagues. That is the control the scheme was built to defeat: a second person looking at the request.
The DOJ lists six victims by type and location, with losses between about $2.7 million and $11.6 million each: a Missouri hospital system, a Rhode Island bank, a California hospital system, a UK fintech, a Massachusetts mutual fund company and a Florida pet food company. Those six don't add up to $129 million. The release gives the total as "more than $129 million" and says the list is "including" those six, so there are other victims it doesn't itemize. One more wrinkle: the DOJ uses the phrase "super intelligence" for the technology throughout. The underlying tools are voice and image synthesis, which is what I mean by AI here.
Why a 20 person company should care
The victims include hospitals and a bank, so you might read this as a big-organization problem. I wouldn't. The attack needs one employee who can send a wire, a believable story, and a reason to keep quiet. A small finance team has all three, and usually fewer checks than a hospital system does.
It also isn't only a voice problem. Voice cloning is the new part. The underlying fraud, a boss asking for a quiet urgent payment, is old, and the FBI's Internet Crime Complaint Center has been warning about it for years under the name business email compromise. What AI changes is that "I'll just call and check" is no longer proof, if the person you call is the fake.
What to put in place
None of this is clever. That's the point. Write it down and make it a rule the CEO is also bound by.
- Call back on a number you already had. Not the number in the email, not the number the caller gives you. Use the one in your own contacts or your HR system. If the CEO is the person asking, call the CEO's known mobile, or walk over.
- Two people approve every wire above a threshold. Pick a number that fits your business. The second approver should not be someone the requester can lean on, and should see the request independently.
- Ban "don't tell anyone." Say out loud in your finance procedure: any payment request that asks you to keep it from colleagues goes straight to a second person. Secrecy is the tell, and the script in this case was written to create it.
- Treat new payees and changed bank details as high risk. Verify them by a separate channel before the first payment, every time, even for a vendor you know.
- Use a code word that isn't on the internet. Anyone with public videos or podcasts has handed over audio to work from. A phrase agreed in person, not written in email or chat, is something a clone doesn't have.
- Tell the bank fast if it goes wrong. A recall request sent the same day has a far better chance than one sent next week. Know the bank's fraud desk number now, and report to the FBI at ic3.gov. The DOJ release also points possible victims to tips.fbi.gov.
Rehearse it once. Run a drill where someone calls your finance person with a plausible request and see whether the process holds. It is uncomfortable, and far cheaper than finding out for real.
Where a pen test does and doesn't help
I run a pen testing business, so I'll say it plainly: an external pen test does nothing about this attack. It looks at what your servers and sites expose to the internet. It doesn't test your people, your phone calls or your payment approval process. That's a process and training fix, and the controls above cost nothing but discipline.
For what an outside test does cover, see what a $495 pen test covers, and for how AI is changing the attacker's side more broadly, AI-powered hacking.