US charges three men with using AI-cloned executive voices to steal $129 million from companies

• The AI used: voice cloning on phone and video calls, plus image or video impersonation of CEOs "sometimes" (CBS News). None of the reviewed coverage names a model, product or vendor.
• The human side: an opening email asking for help with a confidential deal, a live call from the fake CEO, then a handoff to a fake lawyer. Soall allegedly worked from a handwritten script titled "Lawyer Call" (SecurityWeek).
• Getting the money out: accounts in at least nine countries, including the US, Spain, Portugal, Greece, the UAE, Singapore and China, followed by conversion to crypto (Guardian).

The public record contains no technical details: no synthesis method, no indication of real-time or pre-recorded audio, no tools named, and no description of how well the fakes held up on live calls. Three points stand out from what is documented:

• AI was one part of a larger script. The fraud depended on a chain of identities (CEO, then lawyer) and on reusable scripts, which points to a repeatable playbook rather than one-off fakes (SecurityWeek).
• The law-firm handoff passes trust to a second fake person. A check that only confirms the CEO's identity would miss the fraud.
• The scale points to automation. Prosecutors say "thousands" of companies were targeted, but how many were defrauded has not been disclosed (CBS News).

All case facts come from DOJ allegations as reported in the press. None has been tested in court, and no independent forensic analysis of the AI is public.

• Total stolen: about $129m (CBS News). The Guardian's opening line says "over $100m" and its body says $129m (Guardian).
• Individual victims: reported figures differ a little. The Missouri hospital system lost about $5m per the Guardian and about $4.8m per TNW. The UK tech company lost over $11m per the Guardian and about $11.6m per TNW (TNW).
• Prior record: Hazan was convicted of conspiracy to commit bank fraud in New York in 2008 (SecurityWeek).

• Business email compromise (BEC) is when fraudsters pose as executives or partners to redirect payments. It is an old fraud, and AI voice and video add a convincing live layer on top.
• The Arup case showed that a multi-person deepfake video call could work. No arrests were publicly reported there.
• The FBI warned in May 2025 about AI voice deepfakes impersonating senior US officials, a sign that voice cloning had become common (BleepingComputer).
• The new case is unusual because it names alleged perpetrators and attaches a large total loss to them.

• Everything here is an allegation. The defendants are presumed innocent, and all three are still awaiting extradition (CBS News).
• The sources do not say how much of the success came from AI and how much from ordinary social engineering.
• Reported victim amounts differ slightly from outlet to outlet, and the indictment text was not among the reviewed sources.
• Industry figures on deepfake losses that circulate online come mostly from vendors. They do not support conclusions about this case.

• Don't trust a voice or face on a call. Payment approvals should need confirmation through a separate, known channel, such as a call back to a number already on file.
• Check both identities. Confirm the "executive" and the "outside counsel" separately, never using contact details that came with the request.
• Watch for the warning signs. Requests marked confidential, transfers abroad and multi-country payment routes match the pattern alleged here.
• Know which sectors were hit. Hospitals, banks and asset managers were all named victims (Guardian).
• Deepfake detection is not enough on its own. Payment controls work no matter how good the fake is.