Four incidents that defined deepfake fraud.

The deepfake incidents worth studying are the ones with a public record: a police statement, a court filing, a company confirmation, or reporting that named its sources. These four span audio, video, politics and corporate finance, and together they map how the attacks have evolved. We have left out the anonymous cases that circulate with round-number losses and no verifiable detail.

2019. A cloned voice and a Hungarian supplier

In March 2019 the chief executive of a UK energy firm took a call from someone who sounded exactly like the head of its German parent company. The voice asked for an urgent transfer of 220,000 euros, about $243,000, to a Hungarian supplier, and promised it would be reimbursed. The money moved on to Mexico and beyond. A second call promising the reimbursement, and a third from an Austrian number, finally raised suspicion. The firm's insurer, Euler Hermes, later described the case to the Wall Street Journal as the first voice-cloning fraud it had seen.

What it taught: a voice carries authority on its own. The attacker did not need video, documents or system access, only a plausible request, a familiar accent and a reason to hurry.

2022. A surrender that never happened

On 16 March 2022 a roughly one-minute video appeared showing Volodymyr Zelensky telling Ukrainian soldiers to lay down their arms. It was placed on a hacked Ukrainian news site, pushed onto the Ukraine 24 television ticker, and circulated on Facebook, YouTube and Twitter before those platforms removed it. The quality was poor: the accent was off and the head did not sit right on the body. Zelensky posted a rebuttal on Telegram within hours, and Ukraine's government had been warning for weeks that such a video might come.

What it taught: a bad deepfake still buys the attacker a window of hours, and that window is where the damage is done. Pre-warning the audience and having a fast, credible channel to respond shortened the window dramatically.

2024. The New Hampshire robocall

On the eve of the 23 January 2024 presidential primary, voters in New Hampshire received robocalls in a synthetic version of President Biden's voice urging Democrats not to vote. The calls were commissioned by political consultant Steve Kramer and carried by Lingo Telecom. The Federal Communications Commission proposed a $6 million penalty against Kramer and Lingo settled for $1 million. Kramer was indicted in New Hampshire on 13 felony counts of voter suppression and 13 misdemeanor counts of impersonating a candidate, and was acquitted by a jury in June 2025.

What it taught: the technical bar for a convincing voice at scale is now trivial, and the legal response is still uneven. Regulators moved against the carrier and the organizer within months; a criminal jury did not convict. Provenance at the carrier level, not detection after the fact, was what produced consequences.

2024. Fifteen transfers on a video call

Early in 2024 an employee in the Hong Kong office of the engineering firm Arup received a message that appeared to come from the company's UK-based chief financial officer about a confidential transaction. The employee joined a video conference in which the CFO and several colleagues appeared and spoke, then made 15 transfers totalling HK$200 million, roughly $25 million. Hong Kong police disclosed the case in February 2024 and Arup confirmed in May that fake voices and images had been used. No arrests had been reported at the time.

What it taught: a live, multi-person video call is no longer proof of who is on it. Verification has to happen through a channel the attacker does not control, and large transfers need a check that does not depend on what the requester looks or sounds like.

What they have in common

• A request that was plausible on its face, made through a channel the victim already trusted.
• Urgency, confidentiality or both, to discourage the one phone call that would have ended the attack.
• A window between the first contact and the loss in which detection would have changed the outcome.
• Consequences that arrived, if at all, long after the money or the message had moved.

The pattern is why we treat deepfake detection as part of a fraud workflow rather than a standalone media check. The signal has to arrive while the decision is still open, with enough evidence attached that a reviewer can act on it.

All four incidents are reconstructed from public reporting and official statements listed below. Figures are as reported by those sources. No proprietary model internals or customer-specific details are included.