A Video Call Is Not a Payment Approval

Singapore, March 2025: the pattern, step by step

According to Singapore Police, the finance director was contacted on WhatsApp on March 24, 2025 by a scammer impersonating the company’s chief financial officer, and joined a Zoom video conference the next day in which the chief executive, among others, was impersonated using deepfake technology. Before the conference he had been told it concerned a restructuring of the firm’s regional business, had received a call from a purported lawyer stressing confidentiality, and had been made to sign a non-disclosure agreement. On March 26, 2025 he transferred more than US$499,000 from the company’s bank account to a local corporate account, which turned out to be a money-mule account; more than US$494,000 was moved on to accounts in Hong Kong.

The fraud surfaced because the fraudsters asked for more. When a further US$1.4 million was requested on March 27, 2025, he realized it was a scam and alerted the bank, which notified Singapore’s Anti-Scam Centre; on March 28 Hong Kong police withheld the full amount in the Hong Kong accounts.

Two weeks before that payment, in March 2025, Singapore’s police, central bank and cyber-security agency had issued a joint advisory describing a similar method: scammers impersonate senior executives of the victim’s own company on live-streamed Zoom calls reached through unsolicited WhatsApp messages, and instruct transfers from corporate accounts under the pretext of business payments. Victims are passed to a second scammer posing as legal counsel, who sends a non-disclosure agreement or board letter to a personal email address. The advisory says victims realized what had happened when the scammers became uncontactable or when they checked with the real executives and lawyers, who confirmed they had never joined any call or authorized any transfer.

Arup, 2024: the case that set the pattern

Arup’s loss remains the best-documented case of its kind, with a named company that confirmed it. Hong Kong police said the employee attended a video call with what he thought were several other members of staff, all of whom were deepfake recreations. The employee then made 15 transfers totaling HK$200 million to five bank accounts, reported by CNN in February 2024 as about US$25 million. Arup notified Hong Kong police in January 2024, later confirmed that fake voices and images were used, and said none of its internal systems were compromised. Its chief information officer has called the incident “technology-enhanced social engineering”.

The deepfakes did not have to be perfect. In June 2024, Hong Kong’s Secretary for Security told the Legislative Council that in a case reported to police at the end of January 2024, with HK$200 million lost by an employee invited to a video conference by someone impersonating the UK head office’s chief financial officer, police believed the conference was pre-recorded from public video clips and voices, so there was no interaction. The written reply does not name the company; its date, amount and method match the Arup case. After giving the instruction, the fraudster ended the meeting on a pretext and continued the payment instructions by instant messaging. Arup’s chief information officer later tried to make a real-time deepfake of himself with open-source software; it took about 45 minutes and “wasn’t particularly convincing”. The call only had to look like confirmation.

Why a familiar face or voice no longer proves anything

The fakes are now easier to make and harder to spot. The FBI warned in May 2025, and again in December 2025, that AI-generated content is often difficult to identify, and that a call or voice message from a known contact and an AI voice clone can sound nearly identical. In one experiment of a peer-reviewed study published in September 2025, 50 UK listeners labeled voice clones made from about four minutes of recordings as human on 58% of trials and could not tell them from the real voices better than chance. The speakers were strangers to the listeners, and a clone may convince less someone who knows the speaker well. The International AI Safety Report 2026, published in February 2026, says AI-generated content has become harder to distinguish from real media since January 2025, while noting that multimodal outputs combining video, audio and text appear easier to detect than text or audio alone. Europol’s 2025 serious-crime assessment says voice cloning and live video deepfakes are easily accessible, need no specific technical skills, and are being used in CEO fraud.

Voice alone is enough. In January 2026, police in the Swiss canton of Schwyz reported that a company owner made several transfers totaling several million francs to an account in Asia after a caller with an AI-manipulated voice convincingly posed as a known business partner; the fraud was recognized only after the payments were complete.

The numbers are reported complaints, not a measure of how often this happens. The FBI Internet Crime Complaint Center’s annual report for 2025, covering January to December 2025, records more than 22,000 complaints with AI-related information in 2025, with adjusted losses above $893 million, including more than $30 million reported by businesses for business email compromise involving AI; business email compromise complaints as a whole reported losses of more than $3.0 billion. The International AI Safety Report 2026 notes that systematic data on AI-enabled fraud remain limited and that embarrassment or fear can make victims reluctant to report it. In the UK, CEO fraud was under 1% of authorized push payment fraud cases in 2025 but carried the highest average loss of the eight authorized push payment scam types it tracks, just over £28,000 per case, according to UK Finance’s annual report published in June 2026.

Controls for payment requests

• Confirm through a channel you choose. Call the requester back on a number taken from the company directory or vendor master file, never one supplied in the request or the meeting invitation.
• Make secrecy and urgency reasons to slow down. State in policy that no one, at any level, may instruct a payment that bypasses the approval workflow, so that staff can refuse without risk.
• Require a second approver for large payments, new beneficiaries and changed bank details, someone who did not receive the original request.
• Watch for splitting. Several payments to new accounts on one instruction should trigger review, whatever the individual amounts.
• Pre-agree a verification step for executives, such as a callback to an assistant or a confirmation inside the payment system itself, so that confirming a request does not depend on recognizing a face or a voice.
• Train and rehearse. Brief the people who release payments on this pattern, test the procedure with a simulated request from a senior person, and record whether the callback happened.
• Know the recovery path. Keep the bank’s fraud line where payment staff can find it, and make calling it the first step when a payment is in doubt.

Are you at risk?

Answer these for the people in your organization who can release payments.

1. Is a second approver required before one person can send a large payment to a new account on the strength of a call, video meeting or message?
2. Would they call back on a number from your own records before paying, rather than on a number or link the requester supplied?
3. Does your policy say that no one, however senior, may instruct a payment outside the approval workflow, and that a request for secrecy is a reason to stop?
4. Are new beneficiaries and changed bank details verified by someone who did not receive the request?
5. Would several payments to new accounts on one instruction be flagged before they all went out?
6. Have the people who make transfers been briefed on this pattern in the last year, and has the procedure been tested with a simulated request?
7. Does everyone know to call the bank immediately if a payment might be fraudulent?

A “no” to any of the first four identifies a gap in the checks these cases make relevant. Establish whether another documented control covers it, then test that control with a simulated request before relying on it.