Deepfake fraud does not break in. It applies.

On 14 May 2026 the Singapore Police Force published an advisory about scams impersonating senior government officials, including the Secretary to the Cabinet. It updated an alert from 6 May, and it says the quiet part plainly: the scammers “appear to be targeting business professionals who have had prior interactions with government officials.”

In the case the advisory describes, the victim received a WhatsApp message carrying a government official’s profile photo, and an email asking for urgent funding assistance. Attached was what appeared to be an official letter of guarantee, bearing the Prime Minister’s signature, promising reimbursement by the Singapore Government within 15 business days. The victim was asked to endorse a non-disclosure agreement and furnish a copy of his identification card. He did both.

Then came the meeting. A video conference that appeared to involve the Prime Minister and other officials, local and overseas. It did show their faces and carry their voices. The Police state that the participants had been fabricated using deepfake AI technology. After the call, the victim transferred at least SGD 4.9 million through a series of transactions to a corporate bank account the scammers supplied. He realised something was wrong only when he contacted the real Secretary to the Cabinet — on the day the advisory was published.

“The scammers appear to be targeting business professionals who have had prior interactions with government officials.”

Singapore Police Force advisory, 14 May 2026

Nothing in that chain was a breach

Read the sequence again and notice what is missing. No malware. No stolen password. No exposed server. Every step was a decision a competent person made: endorse a document, send an ID, join a call, approve a transfer. The attacker did not break into the process. They applied to it.

That is what changes the defence against deepfake fraud. Training staff to watch for odd lighting or unnatural blinking is worth something, and GovTech Singapore publishes exactly that checklist. But it says the honest thing beside it — that even the most advanced detection systems cannot catch every forgery in real time. A control that depends on a busy person noticing an artefact, on a small video call, against a deadline, is not a control. It is a hope with a training slide.

The advisory ends with four things government officials will never do over email or phone: ask you to transfer money, ask for your bank login details, ask you to install an app from an unofficial store, or transfer your call to the Police. Read those as detection tips and they are weak. Read them as statements about the channel and they are strong, because a real government channel cannot do those things no matter how convincing the video is. That is the shape of a rule that survives contact with a good fake: it does not ask anyone to judge the forgery.

How often this lands on businesses

This is not one unlucky person. A survey fielded in August 2024 by Sapio Research for Regula — 575 business decision-makers across financial services, crypto, technology, telecommunications, aviation, healthcare and law enforcement — found that 56% of Singaporean businesses reported incidents of audio deepfake fraud, and 52% reported video deepfakes. Those are incident reports, not audited losses, and they are from 2024 fieldwork rather than this month.

The same survey’s least glamorous finding is the one worth keeping: 58% of businesses worldwide had encountered fake or modified physical documents — more than the share hit by deepfake fraud. The old fraud did not disappear. It got a better disguise, which is why the defence below is about process rather than about spotting synthetics.

For scale on the synthetic side, the FBI’s 2025 Internet Crime Report carried a section on artificial intelligence for the first time in nearly 25 years of reporting: 22,364 complaints, costing Americans nearly $893 million, inside 1,008,597 complaints and almost $21 billion in reported losses overall. Scammers, the FBI notes, lean on pressure — and now deploy fake profiles, voice clones, identification documents and believable video of public figures or loved ones to apply it.

What AI fraud defence actually is

Strip away the synthetic face and every deepfake fraud attempt is a request for money arriving through a channel that looks trusted and carries its own urgency. The defence is not better eyesight. It is a rule about which channel is allowed to authorise what — and it belongs in the payment process, not in a training deck.

1. Verify out of band. Any request that moves money gets confirmed on a channel the requester did not choose and cannot control — a number you already held, dialled by you. The Police’s “never transfer your call” rule is the same principle, and it holds because it never asks anyone to judge a face.
2. Make the second approver real. Above a threshold, two people approve, and the second is told what to check — not merely what to sign.
3. Write the runbook down. If the payment process lives in one manager’s habits, the attacker’s urgency lands on that manager alone, at the end of a long day.
4. Stop sending identity documents over chat. An ID card sent in a WhatsApp thread is a permanent asset for whoever controls that thread. It was part of this case.
5. Reduce what can be cloned. GovTech’s advice to limit the photos, videos and voice recordings staff publish is the cheapest control here, and the least popular.

The other half of the problem is your own data

The raw material for deepfake fraud is public by default: conference photos, a podcast appearance, a webinar recording, a voice on the company video. Nobody can withdraw all of it. But a business can stop adding to it deliberately — and can know what an AI agent acting on its behalf is permitted to read, say and send. That is the ground we build on in AI Privacy Guard and Agentic Organisation, and it is why we verify money paths against a source system rather than a screenshot in Accounting Verification.