A finance employee joins a video call. The CFO is on screen, along with a few familiar colleagues. The CFO explains a confidential transaction and asks for payments to be made quickly. The employee does what any diligent employee would do and makes the transfers. Every face and voice on that call, except the employee's, was generated by software.
This is not a thought experiment. It is close to what Hong Kong police described in early 2024, in the case later identified as involving the engineering firm Arup. This article explains what happened, why detection tools alone will not save you, and the process controls that protect a business even when nobody on the call can tell the fake from the real person.
What happened in the Hong Kong deepfake video-call fraud?
According to reporting by CFO Dive, an employee in Arup's Hong Kong office received a message that appeared to come from the company's UK-based CFO about a "confidential transaction". The employee then joined a video conference with the fake CFO and other AI-generated staff, and afterwards transferred about HK$200 million (roughly US$25 million) to five Hong Kong bank accounts. The fraud surfaced only when the employee followed up with head office. Arup notified Hong Kong police in January 2024 and later confirmed that fake voices and images were used. Hong Kong police believe the "video conference" was pre-recorded, built from online videos and audio of the real people, which shows how little an attacker needs.
The Hong Kong government's own reply to the Legislative Council in June 2024 lists this case alongside another from May 2024, in which a staff member of a multinational trading company joined a video call of nearly 30 minutes with a fake boss and then transferred nearly HK$4 million. Two cases, two companies, the same pattern: a believable senior person, urgency, secrecy and a payment instruction.
The root cause was a process gap, not a technology gap
It is tempting to read this story as "the deepfake was too good". The more useful reading is that a single person, acting on instructions received through one channel, was able to move a very large sum without an independent check. The deepfake defeated the human eye. It should never have had to defeat a control.
Can deepfake detection tools reliably spot a fake CEO video?
Deepfake detection tools exist and they are improving. They look for artefacts such as inconsistent lighting, unnatural blinking, lip-sync errors, compression patterns and audio anomalies. They are worth using in some situations, for example screening uploaded identity videos in a KYC flow. But a business owner should understand their limits before relying on them:
- It is an arms race. Each new generation of generation tools is trained to remove the artefacts the previous detectors looked for. A detector that worked last year may miss this year's fakes.
- Live calls are hard to scan. Most video-call platforms do not run third-party detection on the stream, and low-bandwidth video hides the very artefacts a detector needs.
- Scores are probabilities. A tool that says "72% likely genuine" does not tell a junior accountant whether to send ₹2 crore.
- Attackers choose the channel. If video is checked, they switch to a voice note, a WhatsApp call or an email "following up on our call".
So treat detection as one signal, not the decision. The decision must rest on a process that works even if detection fails completely.
Manual checks during a suspicious video call
The Hong Kong Police advice, quoted in the government's reply, is practical and costs nothing:
- Ask the person to turn their head up, down, left and right, and watch for distortion around the face.
- Ask a question only the real person would know, ideally about something not visible on social media.
- Confirm any money request made over video or audio with a separate phone call.
- Do not answer unknown video calls.
These checks raise the cost for the attacker, but they are still checks performed by a stressed person in real time. The stronger protection sits in your payment process.
Process controls that work even when detection fails
Below is the control set we recommend to finance teams. None of it needs new software, and most Indian companies already have half of it on paper.
| Control | What it means in practice | Why it beats a deepfake |
|---|---|---|
| Call-back on a known number | Any payment instruction received by call, video, email or chat is confirmed by calling the requester on a number already saved in the company directory, never one supplied in the message. | The attacker controls the inbound channel, not your directory. |
| Dual authorisation | Payments above a set threshold need two approvers, one of whom was not part of the original conversation. | One deceived person cannot complete the fraud alone. |
| New-beneficiary hold | Payments to a bank account added in the last few days are held for a cooling period and verified separately. | Fraud almost always pays into fresh accounts. |
| No secrecy exception | A written rule that "confidential" or "urgent" requests still follow the normal approval path, and that senior staff will never ask anyone to skip it. | Secrecy and urgency are the attacker's main tools. |
| Split-payment alert | The bank or ERP flags a series of transfers to multiple new accounts in a short window. | The Hong Kong case involved several transfers to five accounts. |
| Code phrase for executives | Senior leaders and finance staff agree a verbal challenge phrase, changed periodically and never written in chat. | A cloned face and voice do not know it. |
Make the CEO publicly say it
The control that most often fails is the social one. A junior employee will not challenge the CEO unless the CEO has said, clearly and on record, "I will never ask you to skip the approval process, and I will thank you for calling me back." Put that sentence in the finance policy and repeat it in a town hall. It costs nothing and it removes the fear that makes these frauds work.
Reducing your exposure before an attack
Deepfakes are built from public material. You cannot remove your CEO from the internet, but you can be deliberate:
- Know what is public. Keynote videos, podcasts and investor calls give attackers clean voice and face samples. That is a reason to tighten the payment process, not to stop speaking.
- Separate identity from authority. No one should be able to authorise a payment simply by being recognisable. Authority should come from the approval system, not from a face on a screen.
- Train with real examples. Show staff the Hong Kong case and run a drill: a "CFO" calls and asks for an urgent transfer. Measure who calls back.
- Prepare a reporting path. Staff should know exactly whom to tell, and that reporting a suspected fake will never be punished.
What Indian rules say about deepfakes
India amended the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules in February 2026. As summarised by Freshfields, the amendments, effective 20 February 2026, define "synthetically generated information" to cover audio, visual or audio-visual content that appears real, require such content to be clearly labelled, and shorten takedown timelines, including a two-hour window for certain high-risk content such as impersonation. These duties fall mainly on platforms. They help with removing a fake video of your CEO after it spreads; they do not stop a private video call aimed at your finance team. That part is still your job.
If you think you have been hit
- Call your bank immediately and ask for the beneficiary accounts to be frozen. Speed matters more than anything else.
- In India, report financial cyber fraud on the national cybercrime helpline 1930 and on cybercrime.gov.in, and file a police complaint.
- Preserve evidence: meeting links, call logs, chat messages, email headers and the payment trail. Do not delete the meeting recording or the messages.
- Review whether the attacker had access to internal information, such as an email account, that made the story convincing. A deepfake call is often the second stage of an earlier compromise.
How Cehpoint can help
Payment-verification rules are security governance, and that is where we usually start. Our cyber security team, including a virtual CISO on retainer, can help you write these controls into policy and check whether they hold up in practice. Where an incident has already happened, our cyber-crime investigation and digital forensics team helps preserve evidence and trace how the attacker got in. If you want a quick view of where your payment process is weakest, a paid consultation with a senior engineer is a sensible first step.