Cloned voices in vishing: from the Arup deepfake call to the family code word
On the telephone, a person's voice used to be close to proof of identity. Cheap voice-cloning tools have quietly ended that assumption. A fraudster with a few seconds of audio — a podcast clip, a voicemail greeting, a video posted online — can produce speech that relatives, colleagues and call-center staff cannot reliably tell from the original. The technique feeds vishing — voice phishing — and by 2026 it has documented corporate and family victims, plus research showing why human listeners and some automated checks miss the fakes.
The HK$200 million video call
The case that moved voice cloning from novelty to boardroom risk unfolded in Hong Kong in January 2024. A finance worker at a multinational firm received a message that appeared to come from the company's UK-based chief financial officer, describing a secret transaction. He suspected phishing — until he joined a video conference where the CFO and several colleagues he recognized appeared on screen. According to Hong Kong police, every participant was a deepfake. The employee made 15 transfers totalling HK$200 million — about US$25.6 million — to five local bank accounts, and the fraud surfaced only when he checked with head office. As senior superintendent Baron Chan put it:In the multi-person video conference, it turns out that everyone [he saw] was fake.The engineering firm Arup was later identified as the victim and confirmed the incident in May 2024, saying it had notified police early in the year and that its financial stability, operations and internal systems were unaffected. Its chief information officer, Rob Greig, noted that attacks including invoice fraud, phishing scams, WhatsApp voice spoofing and deepfakes had risen sharply in sophistication. Police classified the case as obtaining property by deception; no arrests had been announced at that point.
Family emergencies with a familiar voice
The same technique scales down to ordinary households. The US Federal Trade Commission warned about it in March 2023: in the classic grandparent scam — a call from a relative who has wrecked the car and landed in jail — the caller can now be a clone. All the scammer needs, the agency wrote, is a short audio clip of a family member's voice, taken from content posted online. The FTC's advice is procedural: hang up, call the person back on a number you know is theirs, check with another relative, and treat requests to wire money, send cryptocurrency or buy gift cards as red flags. European cases show the same pattern against other kinds of targets. In Italy, an elderly woman was defrauded of €30,000 in a call that used what sounded like her daughter's voice, according to Guardian reporting from February 2025. That same month, Milan prosecutors received complaints from prominent business figures — among them staff of Giorgio Armani and Prada chair Patrizio Bertelli — after fraudsters cloned the voice of defence minister Guido Crosetto and phoned them, claiming to raise ransom for kidnapped journalists, with reimbursement promised by the Bank of Italy. At least one victim transferred a total of €1 million to an account in Hong Kong.Why ears, and some checks, fail
Human listeners are a weak line of defense. Pindrop's 2025 Voice Intelligence & Security Report, based on an analysis of more than 1.2 billion customer calls, recorded a rise of more than 1,300 percent in deepfake fraud attempts across contact centers in 2024 — from an average of one attempt per month to seven per day — along with a 173 percent increase in synthetic voice calls between the first and last quarters of the year. Synthetic voice attacks on banks grew 149 percent, and the report projected that contact centers could face roughly $44.5 billion in fraud exposure in 2025. Biometric verification is not a catch-all either: Hong Kong police said deepfakes were used at least 20 times in related fraud to defeat facial recognition checks imitating identity-card holders. Voice-based checks face the same problem: a clone is a synthetic version of the real signal, and short, low-quality telephone audio makes automated detection harder, not easier. Detection vendors are improving, but the caller side of the race is industrialized too, and the volume, by Pindrop's own numbers, is growing faster than the filters.What still works
Given the limits of detection, the defenses that hold up are procedural. Suggestions on the FTC's alert included family code words — a phrase agreed offline that any genuine emergency call must include — and verification questions only the real relative could answer, though such details can leak through social media. Callback discipline on a known, independently dialed number remains the most reliable single step, because it removes the attacker's channel from the loop:- Agree a family code word for emergencies, and use it.
- Hang up and call back on a number you already have — not one supplied by the caller.
- Treat urgency, secrecy and unusual payment rails such as wire transfers, crypto or gift cards as the scam's signature, not a detail of it.
- Limit the clean voice samples you publish in voicemail greetings and public video.