In January 2024, a finance employee in the Hong Kong office of Arup, the British engineering firm behind the Sydney Opera House, joined a video call with the company’s chief financial officer and several colleagues. The request was confidential and urgent. The faces were familiar. So were the voices. Believing them, the employee sent a total of HK$200 million, about US$25.6 million, to the fraudsters.
None of the people on the call were real. Hong Kong police later said every participant had been a digital recreation. When Arup confirmed in May 2024 that it was the company involved, it said that “fake voices and images were used,” and that its financial stability and internal systems had not been affected.
The case is often told as a story about video. It is really a story about trust. According to Hong Kong police, the employee had suspected the first message was a phishing email. What settled the doubt was seeing and hearing people they knew. That instinct, to check by listening, is exactly the one that no longer works.
What changed
The technology did not arrive overnight, but its cost collapsed. In March 2024, OpenAI described a model it had been testing privately, Voice Engine, that could produce natural-sounding speech “that closely resembles the original speaker” from text and “a single 15-second audio sample.” The company chose not to release it widely, citing the risk of misuse, and made an unusual public recommendation: that society begin “phasing out voice based authentication as a security measure for accessing bank accounts and other sensitive information.”
That is a striking sentence for a technology company to publish about its own product category. It amounts to an admission that the voice, as a credential, is finished.
Other tools with similar abilities are widely available. The point is not any single product. It is that a short clip from a podcast, a conference panel, a voicemail greeting or a family video is now raw material.
What the authorities are saying
The clearest signal comes from the agencies that count the losses.
In December 2024, the FBI issued a public service announcement on criminals using generative AI for financial fraud. It listed vocal cloning among the techniques, noting that criminals “generate short audio clips containing a loved one’s voice to impersonate a close relative in a crisis situation,” and “obtain access to bank accounts using AI-generated audio clips of individuals and impersonating them.”
In May 2025, the FBI warned of an ongoing campaign in which malicious actors had been impersonating senior US officials since April of that year, using text messages and “AI-generated voice messages” to build rapport with targets, many of them current or former officials and their contacts. In December 2025 it updated that warning: the activity dated back to at least 2023, and the impersonators had targeted officials’ family members and personal acquaintances. Their requests, the FBI said, included authentication codes, copies of passports, introductions to associates and wire transfers to overseas institutions. Among the lures: an invitation to meet the president, or word that the target was being considered for a company board.
That last detail matters for a wealthy readership. The pitch is not crude. It is flattering, specific and plausible, aimed at people whose diaries really do contain boards, introductions and meetings with senior figures.
One case became public in July 2025. A State Department cable, reported by The Washington Post and CNN, said someone using AI to impersonate Secretary of State Marco Rubio had contacted at least five people, including three foreign ministers, a US governor and a member of Congress, leaving voicemails on Signal “with the goal of gaining access to information or accounts.” The department said it was investigating.
Financial regulators have reached the same conclusion from the other side of the counter. In November 2024, the US Treasury’s Financial Crimes Enforcement Network (FinCEN) issued an alert to banks on fraud schemes involving deepfake media, reporting “an increase in suspicious activity reporting by financial institutions describing the suspected use of deepfake media,” particularly to get around identity checks.
The numbers, read carefully
The FBI’s Internet Crime Complaint Center (IC3) published its 2025 annual report in April 2026. Total reported losses across all cyber-enabled crime were US$20.877 billion, a 26 per cent rise on 2024.
For the first time the report broke out complaints involving artificial intelligence. IC3 received 22,364 complaints that mentioned AI, with adjusted losses of about US$893 million. Within that, it reported more than US$30 million lost to business email compromise involving AI, over US$5 million to “distress” scams in which “voice cloning technology is used to mimic the sound of a loved one,” and more than US$632 million in investment complaints with an AI connection.
Two cautions belong beside those figures. First, they count only what people reported, and only where the report mentioned AI. IC3 itself notes that overall investment-scam losses exceeded US$8 billion, “demonstrating that many victims do not realize the extent AI may be involved.” The true figure is unknown, and is almost certainly higher. Second, these are American complaints to an American agency. They describe a pattern, not a global total.
The age profile is also worth noting. Complainants aged 60 and over reported US$7.7 billion in losses in 2025, the largest of any age group, with an average loss of about US$38,500. The people most likely to hold wealth are also, statistically, the people most often targeted.
The law has moved, partly
Regulators have begun to respond. In February 2024, the US Federal Communications Commission ruled unanimously that calls made with AI-generated voices count as “artificial” under the Telephone Consumer Protection Act, which made voice cloning in robocalls illegal and gave state attorneys general new grounds to pursue the callers. The ruling took effect immediately.
That helps with mass robocalls. It does little against a single, carefully researched voicemail sent to one person on an encrypted app by someone outside the United States. Law tends to follow fraud at a distance, and this kind of fraud moves fast.
What the FBI itself advises
The Risk Edit does not give security advice. But it is worth reporting what the FBI tells the public, because its guidance has a common thread.
In each of the announcements above, the Bureau says much the same thing. Verify a caller by hanging up and contacting the person through a number or channel you already have. Be wary when a known contact appears on a new platform or number. Never share a two-factor code. Consider limiting how much of your voice and image is public. And, in its words, “create a secret word or phrase with your family members to verify their identities.”
Notice what that last piece of guidance implies. The FBI is recommending a shared secret, an idea as old as the password, precisely because the voice can no longer serve as one. Verification is moving from what someone sounds like to what someone knows, and from the channel a request arrives on to a channel both sides already trust.
The quiet shift
Most identity checks people rely on in daily life were never formal. A parent knows their child’s voice. An assistant knows the principal’s. A banker knows a long-standing client. A family office knows how the patriarch phrases an instruction. These informal checks did a great deal of unseen work, and it is the informal checks that synthetic voice defeats most easily. They were never designed to be defended.
That is why the change is easy to miss. Nothing visible breaks. The phone still rings and the voice still sounds right. What has changed is the meaning of the sound.
For anyone whose name, voice and calendar are of interest to others, there is a recorded voice somewhere. The question is no longer whether a convincing copy could be made. It is how a request that sounds right gets confirmed before anyone acts on it.
