Deepfake CEO Fraud Hits $129 Million as Identity Checks Fail
A superseding indictment alleges $129 million in deepfake CEO and lawyer fraud. Voice, face and signature checks now fail, and UK Finance logged £576.4m in APP1 losses.

Why it matters
- A superseding SDNY indictment alleges at least $129 million in payments diverted through AI CEO and attorney impersonations between July 2024 and February 2026.
- An Arup employee paid HK$200 million (about $25 million) in early 2024 after a video call in which the CFO and all other participants were deepfakes.
- Named victims include Missouri hospitals at $4,829,580, a UK FinTech at $11,578,474, and a Florida animal nutrition manufacturer at $5,242,747.
- UK Finance's Annual Fraud Report 2026 records a 19% rise in APP1 fraud losses in 2025, reaching £576.4 million.
- NIST SP 800-63B-4 (2025) allows top assurance levels only with phishing-resistant tokens, while EU Regulation 2018/389 requires dynamic linking for certain payments.
A superseding indictment in the Southern District of New York alleges that three defendants diverted at least $129 million between July 2024 and February 2026 using AI-generated versions of real CEOs and attorneys. The case, United States v. Soall, Hazan and Sheetrit, Superseding Indictment S3 26 Cr. 168, turns synthetic video into a payment credential.
Marcus Bennett's analysis in The Economy, published 17 November 2026, frames the case as a structural break for payment security. Cheap imitation of voice and face now erodes the identity checks banks and companies use to release money, shifting verification costs and delays onto firms least able to absorb them.
Around the beginning of 2024, an employee of Arup paid HK$200 million, roughly $25 million, after a video call with the company's CFO “and others.” Every other participant on that call was a deepfake.
Jabotinsky's 31 December 2025 analysis for EU Law Enforcement, “When seeing isn't believing: deepfakes, financial markets, and the limits of EU enforcement,” documents the case. The employee took the prudent course of double-checking with several apparent senior managers. Every confirmation was synthetic.
The comparison to television is deliberate. In September 2019, series 1 of The Capture showed a police officer watching CCTV of an offence he knew he had not committed, with “Correction” footage inserted throughout. Reviewers at the time dismissed the premise as absurd fantasy. Bennett's point is that financial authentication now carries the same defect, except a fabricated face usually talks on screen, and funds move before verification completes.
How did the fraudsters stage the calls?
The indictment describes a repeating four-step script, as Bennett summarises it in The Economy:
- A target receives an email apparently from the CEO, requesting an electronic signature.
- The target signs a stand-alone confidentiality agreement.
- A call arrives from a purported lawyer at a “major” law firm.
- The lawyer cites securities law constraints and tells the target not to inform colleagues, while a synthetic CEO speaks, sometimes on camera.
Associated Press reported the charges on 6 November 2026 under the headline “3 charged in scheme to use AI to impersonate CEOs and attorneys to steal $129 million.” The three defendants face extradition proceedings, and every allegation remains untested in court.
How much did each victim lose?
- A group of hospitals in Missouri: $4,829,580.
- A UK FinTech: $11,578,474.
- A Florida-based manufacturer of animal nutrition: $5,242,747.
- A major New York investment bank: an attempted payment of about $2.85 million in October 2024, halted before release.
The bank payment is the only documented stop in the sequence, The Economy notes. The indictment gives no explanation for how back-office processing caught it, which leaves that control unexplained rather than proven. Kancelaria Prawna Skarbiec reached the same conclusion in its 7 November 2026 review, “Deepfake C.E.O. fraud: inside the $129 million indictment.”
Video-call impersonation of executives predates generative AI. Between 2015 and 2017, an impersonator appeared on numerous calls wearing a mask resembling Jean-Yves Le Drian, then the French Minister of Defence, and took more than €50 million, according to The Economy.
That effort required a skilled mask-maker, suitable premises and audacity. A short sample of suitable speech now replaces all three, which helps explain why the indictment refers to thousands of attempted victims rather than a single successful deception.
Can identity checks still separate a real executive from a fake?
No attribute survives, in Bennett's assessment. Tone of voice, appearance, signature and knowledge of previous discussions can all be generated from public material, including results calls and other public events. Those traits remain present, but they stop distinguishing one person from another.
Stacking such checks adds little reassurance, because each can draw on the same public sources. Virtually any attribute open to assessment can be captured as input to a generating algorithm, Bennett writes, which leaves few traits on which a payment release can still rest.
The one attribute he treats as reliably secret is cryptographic. The sender holds a secret on a token and produces a fresh response to a variable prompt, so any recorded exchange becomes worthless.
What do the standards require?
NIST's Special Publication 800-63B-4, “Digital Identity Guidelines: Authentication and Authenticator Management,” published in 2025, rejects the premise that any characteristic of a person can be treated as reliably secret. It permits the highest assurance levels only with tokens that resist phishing.
For some EU payments, Commission Delegated Regulation (EU) 2018/389 adds dynamic linking, a control that binds the generated code to a particular payment amount and recipient. A code issued for one transfer then fails against another.
Cryptography proves possession, not consent. It certifies issuance from an appropriate token and remains blind to any preceding deception or duress, Bennett notes. A deceived employee who follows every rule still authenticates correctly.
UK Finance's Annual Fraud Report 2026 supplies the scale. APP1 fraud losses, a UK faster-payment category tracked separately, rose 19% in 2025 to £576.4 million, in payments that were, in The Economy's words, “presumably duly ‘authenticated.’”
Can payment controls absorb the shock?
Goodhart and Vu, writing in VoxEU in 2026 under the title “Trust beyond face-to-face contact: AI and the future of finance,” analysed data from 1981 to 2022 spanning 97 economies. They conclude that the availability and quality of information determine trust in domestic institutions, and model an economy “consisting purely of promises as to future deliveries.”
Higher trust leans toward face-to-face contact, yet they identify no mechanism for sustaining that trust across distance. That leaves procedures rather than biometrics as the practical answer, and Bennett lists the options available to companies:
- Call back on a pre-agreed number.
- Require multiple approvers.
- Cap the amount in any single instance.
- Bind codes to specific amounts and recipients through dynamic linking.
- Screen recipients for anomalies.
Each consumes more staff time, more approvers and more devices. None stops duress, internal complicity or compromised equipment. None catches a well-intentioned but mistaken instruction from an authorised person.
Who pays for the new controls?
Distribution decides. The New York bank in the indictment probably employs serious resources for reviewing outgoing payments, Bennett suggests. The owner of the Florida animal nutrition manufacturer, who lost $5,242,747, may employ a handful of staff, none tasked with challenging instructions that appear to come from the CEO.
Costs therefore rise fastest for payments to unknown third parties, cross-border transfers, remote account opening and day-to-day cash management. That reverses a fintech priority pursued over the last 20 years: reducing the cost of paying remotely now meets an expensive requirement to prove the counterparty is real.
Reporting gaps compound the problem. The FBI's Internet Crime Report 2025, cited by The Economy, records that losses grew far faster than complaints between 2020 and 2025, so the complaint count no longer tracks the money lost.
What happens next?
Most payment activity will continue, Bennett expects, with some of the benefit that lower friction previously delivered. The harder problem is finding a safeguard that satisfies institutions, counterparties and customers once trust moves from a recognisable face to an institutional process.
Any such safeguard must survive a convincing impersonator instructing staff to ignore normal procedure, The Economy notes. One bank apparently managed that in October 2024, though the indictment leaves the mechanism unexplained. Until institutions document controls that withstand synthetic voice and face, the $129 million case marks measured exposure rather than a solved problem.
Source: economy.ac
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
718 articles
Related articles
- Anthropic's S-1, a Claude Agent in the Pentagon, and Midterm ICE Fears
- Adactin maps six fraud defences for Australian businesses facing AI scams
- Judge Sets January Trial in Former OpenAI Engineer Bugging Case
- OpenAI Bans ChatGPT Accounts Behind Fake FBI Recovery Scam
- Australia Warns AI Will Industrialise Scams as Losses Hit $2.18 Billion