In early 2024, a finance employee at Arup — the British engineering firm behind the Sydney Opera House — joined a video call with what appeared to be the company's CFO and several colleagues. The request: a series of confidential transfers, framed with urgency. The employee had initially been suspicious of the email that set up the call; the video call dissolved that suspicion, because everyone on it looked and sounded right. Roughly $25 million was wired across multiple transfers. Every other participant on the call had been a deepfake.
The anatomy
- The hook arrived by email, styled as a confidential matter from the CFO — on its own, it raised doubt.
- The video call was the trust mechanism: familiar faces and voices, reconstructed from publicly available footage, defeated the instinct that had correctly flagged the email.
- Urgency and confidentiality suppressed the one control that would have worked — checking with someone outside the call.
- Multiple transfers extended the window before discovery; the fraud surfaced only when the employee later checked with head office.
Why the controls failed
The failure was not gullibility; it was architecture. The employee applied exactly the verification the situation seemed to allow — visual confirmation of colleagues — and that channel was the one the attackers controlled.
Every control that depended on the employee's judgment inside the interaction failed; the control that survives this attack lives outside the interaction: a callback to a registered number, or a hold on the payment until out-of-band verification is logged, backed by dual authorization.
Note also the sub-threshold logic generalizes: splitting the amount across transfers is the same evasion pattern as sizing a request under a dual-approval limit, the core move in CEO fraud.
What would have stopped it
- An enforced rule that no new-destination transfer releases without a verified callback, regardless of who requested it on what channel.
- A no-exception culture where confidential requests explicitly do not suspend verification.
- Real-time synthetic-media scoring on the call as an early warning — useful, but the enforcement layer is what holds when detection is fooled.