Guide Cybersecurity AI

Deepfakes and identity fraud: how to protect your business

In a case reported to Hong Kong police at the end of January 2024 and made public that February, an Arup employee transferred about US$25 million after a video conference built from deepfaked clips of senior colleagues. Voice cloning and real-time face-swapping are now cheap and widely available. This guide explains how the attacks work, why familiar checks fail and which controls hold up.

11 June 2026
11 min read
Updated 18 November 2026
Key takeaways
  • Some voice cloning tools can produce a recognisable clone from a few seconds of public audio. Longer samples give a better result, and real-time tools can carry a live phone conversation.
  • In the Arup case, reported to Hong Kong police at the end of January 2024, an employee transferred HK$200 million (about US$25 million) after a pre-recorded deepfake video conference.
  • The main control is independent verification: payments and account changes go through an authenticated workflow, and any call-back uses a number registered in advance, never one supplied in the request.
  • Corporate email is not an independent channel. An attacker who runs a deepfake call often controls or spoofs the mailbox that sent the request.
  • A shared code word helps as an extra check, but it can leak and people forget it. Treat it as a complement to dual approval, not a replacement.
  • Remote identity checks need to detect two different attacks: presentation attacks against the camera and injection attacks that feed synthetic video straight into the stream.

The Arup video call

At the end of January 2024, Hong Kong police received a report that an employee in the Hong Kong office of a multinational firm, later confirmed as the engineering group Arup, had received a message purportedly from the company's UK chief financial officer asking for a confidential transaction. The employee was sceptical, and was then invited to a video call with the CFO and several colleagues. Police made the case public in February.

According to the Hong Kong government's account to the Legislative Council, the video conference was pre-recorded, built from public video clips and the voice of the impersonated officer, and there was no real interaction between the employee and the fraudsters. Over 15 transactions the employee transferred HK$200 million, about US$25 million, to five bank accounts. The fraud came to light when the employee later checked with head office. Arup confirmed the incident but has not published technical details of how the call was produced, so treat specific claims about the tools used with caution.

The case matters because the employee did what most awareness training asks: they doubted the email and looked for confirmation. The confirmation step itself was the part the attackers had prepared.

Important context

Deepfake fraud does not need a nation-state attacker. Voice cloning and face-swapping tools are sold as low-cost subscriptions and published in open-source repositories. The FBI warned in December 2024 that criminals use generative AI, including cloned voices and video, to make financial fraud more convincing.

How voice cloning works

A voice cloning model learns the acoustic features of a speaker from a sample of their audio: timbre, rhythm, pitch range and speaking style. Some commercial tools produce a recognisable clone from a few seconds of speech; a minute or more gives a noticeably better result. The sample can come from a LinkedIn video, a webinar, a podcast, an earnings call or a voicemail greeting.

Attackers use the clone in two ways. Pre-recorded clips go into voicemails and voice notes. Real-time use takes one of two forms: text-to-speech, where the attacker types and the model speaks in the target's voice, or speech-to-speech conversion, where the attacker talks and the model changes their voice into the target's as they speak. Speech-to-speech keeps the attacker's own timing and reactions, which makes a live call much harder to spot.

A few seconds
Of audio can be enough for some tools to produce a recognisable voice clone
HK$200m
About US$25 million, transferred in the Arup case reported by Hong Kong police in 2024
2 approvals
Recommended control: two independent approvers in the payment system for any transfer above your threshold

Attack types in the corporate environment

📞
Vishing: AI voice impersonation over the phone
An attacker clones the voice of a CEO, CFO or IT manager and calls an employee with an urgent request: approve a transfer, read out a one-time code or grant system access. The caller knows names and projects from public sources. Finance teams, IT helpdesks and executive assistants are the main targets because they can act quickly on a verbal request. Our guide to voice deepfake CEO fraud covers this payment scenario in more depth.
📹
Deepfaked video calls to confirm a payment
Attackers either play pre-recorded deepfake clips into the call, as in the Arup case, or use face-swapping software to replace their own face with the target's live, usually combined with a cloned voice. The call is used as the confirmation step, often after an email request, so it arrives exactly when the employee is looking for reassurance. Group calls help the attacker: the victim sees several familiar faces and the impostors can keep their cameras still and their contributions short.
📋
Bypassing remote identity checks
Remote onboarding and account recovery often rely on a selfie video matched to an identity document. Attackers combine forged or stolen documents with synthetic faces to open bank and fintech accounts, apply for credit or pass employee onboarding under a false identity. The next section explains the two ways they get the synthetic face past the check.
🔒
Credential resets through the IT helpdesk
An attacker using a known employee's cloned voice calls the helpdesk claiming to be locked out. They give the name, department and enough context from LinkedIn or earlier reconnaissance to pass verbal checks. The helpdesk resets the password or registers a new MFA device controlled by the attacker, who then has access without deploying any malware.

The Arup employee was right to ask for confirmation. The attack worked because the attackers had prepared the confirmation step as well.

Why standard verification fails

Informal verification of a verbal request usually rests on three things: you recognise the voice, you can see the person on video and the caller knows things only an insider would know. A prepared attacker can now fake all three. Phone audio compression hides many of the differences between a real voice and a good clone. Caller ID is easy to spoof, and network caller authentication such as STIR/SHAKEN is not in place on every route. Inside knowledge is often on LinkedIn, the company website or in recorded presentations.

Two weaknesses deserve particular attention. First, corporate email is not an independent channel. Deepfake calls often follow a compromised or spoofed mailbox, so "confirm by email" may send the confirmation straight to the attacker. Second, remote identity checks face two different attacks, which NIST SP 800-63A-4 treats separately:

  • Presentation attacks: the attacker shows the camera a printed photo, a mask, a screen replaying video or a live face-swap displayed on another device. Presentation attack detection (PAD) looks for these.
  • Injection attacks: the attacker bypasses the camera, using a virtual camera, an emulator or a tampered app to feed synthetic video directly into the data stream. PAD alone does not catch this; the check also needs device and stream integrity controls.

Warning signs during a call

Train your finance and helpdesk staff to stop when they notice these patterns:

  • Unusual urgency: the requester says there is no time to verify and the normal process will lose the deal.
  • Secrecy: instructions not to discuss the request with colleagues or legal.
  • A parallel channel: an email with bank details arrives during or right after the call, often from a lookalike domain.
  • Refusing a call-back: the person is "in a meeting" or "has no signal" and will not let you call them back.
  • Odd audio or video: unnatural pauses, flat emotion, a still camera or reluctance to turn the head. Good fakes may show none of these, so absence of signs proves nothing.

Controls that actually work

Independent verification through authenticated workflows

The core control is procedural. Never act on a phone or video request alone. Payments, bank detail changes, credential resets and access grants go through an authenticated system (the payment platform, the ERP or the service desk tool) where the requester and the approver log in with their own accounts and MFA. If you need to speak to the requester, hang up and call them back on a number taken from the HR record or the supplier master file, never one given in the call, the email or the meeting invite.

Dual approval with segregation of duties

Remove verbal authorisation as a valid trigger for any transfer above a defined threshold. Above that threshold, require two approvals in the payment system from people who did not raise the request, and make sure no single person can create a beneficiary and approve a payment to it. This is a process change rather than a technology purchase, and it removes the mechanism that deepfake calls rely on: one person persuaded to act alone.

Separate controls for new beneficiaries and bank detail changes

Treat any new beneficiary or change of bank details as a high-risk event with its own procedure. Confirm the change with a known contact at the supplier using contact details already on file, have a second person approve the change in the system, and consider a holding period or a small test payment before the first full payment to the new account. Many deepfake and email frauds end with a request to pay a "new" account, so this one control catches a large share of them.

Helpdesk identity recovery

A voice that sounds right is not proof of identity for a password reset or a new MFA device. Require a stronger route: a call-back to the number on the HR record, confirmation from the employee's line manager through the ticketing system, or identity verification in person or through a proofing service. Notify the user on their existing channels whenever a reset or new device registration happens, so a genuine employee can raise the alarm quickly.

Code words as a complementary check

A code word agreed in person and never written in email or chat can add a useful extra check for executives and their assistants. It is a weak control on its own: code words leak, get reused and are forgotten under pressure. Use one in addition to the workflow and call-back controls above, never instead of them.

Scenario-based staff training

Generic phishing awareness does not prepare staff for a familiar voice asking for something wrong. Run targeted exercises: play a cloned voice of a senior executive making a fraudulent request and let staff practise the verification procedure. Focus on the most targeted roles: finance, the IT helpdesk and executive assistants. Make clear that following the procedure is never a disciplinary risk, even when the request turns out to be genuine.

Remote identity checks and detection tools

If you onboard customers or staff remotely, ask your identity verification provider how it detects both presentation and injection attacks, how it checks document authenticity beyond image analysis and which device and session signals it uses. Ask for independent test results against current attack methods. Synthetic voice and video detectors look for artefacts of synthesis; some generalise to generators they were not trained on, but accuracy tends to fall against new models and over compressed phone audio. Treat them as an additional layer. A call-back to a pre-registered number works however good the clone is.

AttackMain defenceLimitation
Cloned voice asks for an urgent paymentPayment only through the workflow, with dual approval and a call-back to a pre-registered numberThe call-back number must come from your own records, not from the request
Deepfaked video call confirms a requestNo approval given on the call; confirmation inside the authenticated systemSlows genuinely urgent payments, so it needs visible backing from senior management
Request to change a supplier's bank detailsSeparate change procedure: known contact, second approver, holding period or test paymentFails if the supplier's own mailbox or phone line is compromised
Helpdesk call to reset a password or MFACall-back to the HR number, manager confirmation or identity proofing, plus user notificationAdds friction for genuine lockouts
Synthetic face in remote onboardingPresentation and injection attack detection, document checks, device signalsDetection rates vary between providers and drop against new tools

If a suspicious transfer has already been made

Speed matters more than anything else. Call your bank straight away on the number on its website or your card, explain that the payment was fraudulent and ask it to try to recall the funds and contact the receiving bank. There is no guarantee of recovery: money is often moved on within hours.

  • Reimbursement: the UK mandatory reimbursement rules for authorised push payment (APP) fraud, in force since 7 October 2024, cover consumers, microenterprises and charities, up to £85,000 per claim. Most larger businesses are outside that scheme and depend on their bank's own decision and any insurance cover.
  • Report to the police: in England, Wales and Northern Ireland, report through Report Fraud, which replaced Action Fraud in December 2025. In Scotland, call Police Scotland on 101.
  • Preserve evidence: call logs, recordings, meeting invites and links, emails with full headers and the payment records.
  • Personal data: if personal data was exposed, for example because a mailbox was compromised, assess whether the breach must be reported to the ICO within 72 hours under UK GDPR. Report if it is likely to result in a risk to individuals, and record the breach and your decision either way.

Sources

Ryland Deakin
About the author
Lead Consultant, Cyvra · CISM · CompTIA Security+ · MCP

Ryland has delivered cybersecurity, compliance, and IT management programmes for regulated organisations across the UK and the Netherlands for over 20 years. View full profile

Common questions

How little audio does it take to clone someone's voice?

Some commercial tools can produce a recognisable clone from a few seconds of speech. Longer samples, from 30 seconds to several minutes, capture more of the speaker's natural variation and give a more convincing result. The audio can come from a conference recording, a podcast, a company video or a voicemail greeting.

Can deepfake video work on a live video call?

Yes. Real-time face-swapping tools run on consumer hardware and can be convincing enough for someone who is not looking for manipulation. The best-known case did not even need that. According to the Hong Kong government, the Arup video conference was pre-recorded and there was no real interaction, yet an employee transferred about US$25 million after it.

What is a vishing attack?

Vishing is voice phishing: a phone-based social engineering attack in which the caller impersonates a trusted person or organisation to get credentials, a payment or access. Voice cloning lets attackers impersonate a specific person the target knows, rather than an anonymous caller claiming to be from a bank or the IT department.

How do businesses verify identity without relying on voice or video?

Move sensitive actions into authenticated workflows. Payments, bank detail changes and credential resets are requested and approved inside systems where each person logs in with their own account and MFA, with two independent approvers above a threshold. When a call-back is needed, use a number from your own records, never one supplied in the request. Email from the corporate account does not count as a separate channel, because the attacker may control or spoof that mailbox. A code word agreed in person can be added as an extra check, but not as the main control.

Social Engineering Assessment

Test your team before an attacker does

We run targeted vishing and deepfake simulation exercises for finance teams, IT helpdesks and executive staff, and help you build verification procedures that hold up under real attack conditions.