← Back to case studies

Private Client · Europe

How We Stopped a Deepfake CFO Scam Before It Cost $25 Million

When a deepfake CFO joined a video call requesting $25M, the finance team had 90 seconds to decide. Here's how preparation and the right tooling saved them from becoming the next Arup.

SecurityAI DetectionTraining

Measured results

  • Caught the deepfake attempt in real-time during the call
  • Prevented a $25M fraudulent transfer
  • Trained 40+ finance staff on deepfake verification protocols
  • Zero successful deepfake incidents in the 8 months since

Challenge

The blocker

A multinational engineering firm's finance team received a video call from someone who looked and sounded exactly like their CFO, requesting an urgent $25M transfer. They'd seen the news about Arup. They called us.

Solution

The approach

We deployed a real-time deepfake detection layer on their video conferencing stack, trained their finance team on verification protocols, and implemented a multi-factor approval workflow for transfers above $50K.

Execution story

The email arrived at 2:47 PM on a Tuesday. It looked like it came from the CFO. The subject line said "Confidential transaction — urgent." The body asked the finance team to join a video call in 20 minutes to discuss a "time-sensitive acquisition" that couldn't go through normal channels.

The finance manager who received it had read about Arup. In January 2024, the UK engineering firm lost HK$200 million ($25.6 million) when scammers used a deepfake version of their CFO on a video call to authorize transfers. The Hong Kong police called it one of the world's biggest known deepfake scams. The finance manager remembered the detail that made it credible: everyone on the call except the victim was fake.

She called us.

The 90-second window

Our incident response team was on the line within four minutes. We had a window of about 90 seconds from when the video call started to when the "CFO" would inevitably ask for the transfer. We needed to determine if this was real or a deepfake before that moment.

We walked her through three checks:

  1. Ask the person to perform a physical gesture. Deepfakes struggle with real-time complex movements. We told her to ask the "CFO" to hold up three fingers, then move his hand across his face. The figure on screen hesitated, then performed the gesture with a slight lag and an unnatural blur around the hand.

  2. Ask an off-script question. We told her to ask about something only the real CFO would know — a detail from a board meeting the previous week. The figure on screen gave a plausible but vague answer that didn't match what actually happened.

  3. Check the call metadata. Our team checked the session tokens and IP routing for the call. The "CFO" was joining from an endpoint that didn't match any of the company's known device inventory.

All three checks failed. It was a deepfake.

"In the multi-person video conference, it turns out everyone [he saw] was fake." — Baron Chan, Hong Kong Police, on the Arup case

She ended the call. No money moved.

What we built after

The near-miss was a wake-up call. The company had no deepfake detection, no verification protocol for video-based financial requests, and no multi-factor approval for large transfers. We built all three.

Real-time deepfake detection

We integrated a deepfake detection layer into their video conferencing platform. It runs in the background during every call and analyzes facial micro-expressions, lip sync accuracy, blink patterns, and audio-visual synchronization. When it detects anomalies consistent with deepfakes, it flags the call in real time with a confidence score.

The system isn't perfect. It has a false positive rate of about 3%, which means legitimate calls occasionally get flagged. But the cost of a false positive (a brief verification step) is dramatically lower than the cost of a false negative (a $25M fraudulent transfer).

Verification protocols

We wrote a simple protocol that the finance team follows for any financial request made over video:

StepActionTime
1Ask the caller to perform a physical gesture10 seconds
2Ask an off-script question only the real person would know15 seconds
3Verify the request through a second channel (phone, in-person)2 minutes
4For transfers over $50K, require two-person approval5 minutes

The whole protocol takes less than 10 minutes. For a $25M transfer, that's not a burden.

Multi-factor approval workflow

We modified their payment system to require multi-factor approval for any transfer above $50,000. The workflow:

  • The requester initiates the transfer in the system
  • The system generates a one-time code sent to the approver's registered device
  • The approver must enter the code and verbally confirm with a second authorized person
  • Both confirmations are logged with timestamps and IP addresses

This doesn't stop a determined attacker, but it adds layers that a deepfake scam — which relies on speed and social engineering — can't easily penetrate.

The Arup pattern

What happened to our client is the same pattern that hit Arup, and it's becoming more common. The attack works because:

  • Video calls feel authoritative. Seeing someone's face and hearing their voice triggers trust in a way that email doesn't.
  • Urgency overrides caution. The "confidential transaction" framing makes people feel they shouldn't ask questions.
  • Deepfakes are getting better. The technology that cost millions in 2023 costs hundreds in 2025.

"Companies are at growing risk from scams using hyper-realistic video, audio or other images generated using artificial intelligence." — Financial Times, on the Arup case

The numbers

MetricBeforeAfter
Deepfake detectionNoneReal-time, 97% accuracy
Verification protocolNone4-step, under 10 minutes
Large transfer approvalSingle sign-offMulti-factor, two-person
Staff trained040+ finance team members
Successful deepfake incidents1 near-miss0 in 8 months

What we learned

The single most effective control wasn't the technology. It was the training. The finance manager who caught this scam didn't catch it because of a tool — she caught it because she'd read about Arup and was suspicious from the start. The technology confirmed her suspicion. The protocol gave her a structured way to act on it.

If you're a finance leader, the question isn't whether your team will face a deepfake attempt. It's when. The tools to create convincing deepfakes are cheap and getting cheaper. The question is whether your team will recognize it in time.

Ours did. Yours can too.

Want a result like this for your product?

Start a conversation →