Three men have been indicted in New York for allegedly using AI to impersonate chief executives and law firm attorneys, and to talk finance staff into wiring about $129 million to accounts the fraudsters controlled. The Justice Department says the scheme ran from 2024 into 2026. The defendants have not been convicted, and extradition is pending. This is one of the largest publicly charged cases of AI-assisted business fraud so far, and it matters because the attack is cheap to copy and the defense is a matter of process, not technology.
TL;DR: the case at a glance
| Question | Short answer |
|---|---|
| Who is charged? | Jacob Soall (41, Spanish), Sharon Hazan (54, Israeli), Guy Sheetrit (42, Israeli), per The Record |
| How much? | About $129 million, per prosecutors |
| What is the AI angle? | AI used to spoof executives' voices and images, and to pose as lawyers |
| Which AI tools? | Not named in the reporting we read |
| Where are they? | Soall and Hazan arrested in Spain, Sheetrit detained in the UK; extradition requests pending |
| Status | Indicted. Allegations only |
Agents that act on your behalf need rules. See how explainx helps secure your agents →
What the indictment alleges
The Justice Department's Southern District of New York announcement is the primary source for the charges. We could not load the full text of that page in an automated fetch, so the details below follow The Record's detailed report and NBC News coverage of it. Where outlets differ, we say so.
According to The Record, the men posed as company CEOs or lawyers and contacted lower-level employees. They said a confidential financial transaction was underway and told staff to send large payments to accounts the group controlled. On calls, they allegedly mimicked executives' voices or images.
To add credibility, employees were then directed to talk to supposed lawyers at a large international law firm. Those lawyers were also fake. Prosecutors say AI made the impostors look and sound like the attorneys, and that Soall played the lawyers using a script. Hazan allegedly handled laundering the funds.
U.S. Attorney Jamie McDonald said the three used AI "to impersonate real people and deceive corporate executives," as quoted by The Record. An FBI official, James Barnacle, said criminals "continue to weaponize emerging technology like artificial intelligence," and asked other victim companies to come forward.
The victims and the money trail
The Record lists several victims and amounts:
- A Missouri hospital: nearly $5 million
- A Rhode Island bank: $2.6 million
- A California hospital: $2.9 million
- A UK tech company: $11.5 million
- A Massachusetts finance company: $4.6 million
- A Florida pet food company: $5.2 million
Other outlets round the hospital and bank figures slightly differently, which is common when an indictment gives approximate numbers. The point is the spread: hospitals, a bank, and mid-sized firms across sectors. This was not aimed at one industry.
Funds were allegedly wired to accounts in the US, Europe, the UAE and China, then moved into cryptocurrency. The Record also reports that Hazan was previously convicted of bank fraud in 2008 and reportedly boasted of laundering $400 million since 1994. Treat that boast as reported, not as established fact.
Why the two-layer trick works
The scheme is a modern form of what security teams call business email compromise, with an AI layer on top. Two things make it effective.
First, a trusted voice. A finance clerk who hears what sounds like the CEO is far less likely to question an urgent request. Cheap voice cloning needs only short audio samples, and executives appear in earnings calls, podcasts and conference videos.
Second, a fake authority chain. Sending the employee to talk to a "lawyer" is a social proof step. The employee now has two apparent confirmations of the same story, and both come from the attacker. Verification that stays inside the attacker's channel is not verification.
We have seen the same pattern of AI making fake evidence cheap in other stories on explainx.ai. AI agents were caught faking hand-drawn timelapse proof, and a State Farm lawyer was fined $999 for AI-fabricated citations. In each case the lesson is the same: a convincing artifact is no longer proof of anything.
Call center headset illustration showing a voice channel that can carry an AI impersonation fraud request
What is confirmed and what is not
| Claim | Status |
|---|---|
| Three men were indicted in the Southern District of New York | Reported by DOJ and multiple outlets |
| Loss of about $129 million | Prosecutors' figure; The Record also describes it as more than $100 million |
| AI was used to spoof voices and images | Alleged by prosecutors; specific tools not named in the reporting we read |
| Defendants are guilty | Not established. They have not been convicted |
| Extradition completed | No. Requests are pending |
One detail worth noting: the press release headline in the DOJ URL uses the phrase "super intelligence". The reporting we read does not explain what that means in the case, and it should not be read as evidence of an advanced AI system. We could not load the release text, so we do not rely on the phrase.
How to defend your company
No tool in this story needed to be sophisticated. The controls that stop it are procedural.
- Call back on a known number. For any payment request that arrives by call, video or message, hang up and call the person on a number from your own directory.
- Require two approvers for large wires, new payees and changes to bank details. One person talking to one "executive" should never be enough.
- Treat secrecy as a red flag. "Confidential deal, tell no one" is the script in this case and in most business email compromise cases.
- Verify outside counsel independently. If a "law firm" is introduced mid-request, find the firm's number yourself and ask whether the matter exists.
- Set a shared code phrase for executives and finance staff for out-of-band confirmation of unusual requests.
- Report fast. Quick contact with your bank and law enforcement can sometimes freeze a wire. The FBI is asking other victims to come forward in this case.
Teams that run AI agents with payment or email access face a related risk: an agent can be talked into the same bad action as a clerk. AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions such as unapproved transfers. For the defensive tooling side, see our guide to an open-source AI security operations center, and for how labs are building defensive models, Microsoft's MAI-Cyber-1-Flash.
Agent identity badge on a lanyard, a metaphor for verifying who is really on the other end of a request
What this means for AI policy
Cases like this feed the debate over how to label and trace synthetic media. Detection is an arms race, and provenance labels do not help on a live phone call. That is why regulators and prosecutors increasingly focus on the fraud itself: wire fraud and aggravated identity theft statutes already cover AI-assisted impersonation, and the maximum sentences in this case, reported at up to 104 years combined, show prosecutors using existing law. For another example of how fake AI artifacts reach real institutions, see our coverage of an alleged fake Anthropic Mythos checkpoint on Hugging Face.
What finance and IT teams can do this week
Start with a one-hour tabletop exercise. Ask your accounts payable team what they would do if the CEO called, sounded stressed, and asked for a confidential wire within the hour. Most teams discover that nobody has the authority to say no to an executive, and that the written policy lets one person release funds alone.
Then fix the gaps the exercise exposes. Give every clerk explicit permission to pause a request and verify it, with no penalty if the request turns out to be real. Publish a short list of who may approve what amount. Keep a directory of call-back numbers that does not rely on caller ID or the contact details inside an email thread.
Finally, review what executives publish. Public audio and video of leaders is the raw material for voice cloning. You cannot remove it, but you can make sure that being heard is never treated as proof of identity. That change in habit protects the company against this case's technique and against whatever tool the next group uses.
What to watch next
- Whether the extraditions from Spain and the UK go through, and when the defendants first appear in a US court.
- Whether the government names the AI tools used, which would show how accessible the attack is.
- Whether more victims come forward after the FBI's appeal, which could raise the loss figure.
- Whether banks add AI-impersonation checks for wire requests.
We will add an update section here if prosecutors publish more detail.
FAQ recap
The FAQ above covers the charges, the amounts, the tools, the defenses and the custody status. In short: the case is an allegation, the technique is old fraud with new voice and video tools, and the fix is verification outside the channel the request came in on.
Related reading
- AI agents faking hand-drawn timelapse proof
- State Farm lawyer fined $999 for AI fake citations
- Open-source AI security operations center guide
- Microsoft MAI-Cyber-1-Flash
- Alleged fake Anthropic Mythos checkpoint on Hugging Face
- DOJ press release (SDNY)
Details are accurate as of November 6, 2026. The indictment is an accusation, and the defendants are presumed innocent until proven guilty.