The control failed before the transfer
Reports describing the Fideuram incident allege that attackers combined a cloned executive voice with a fake WhatsApp identity to pressure staff into high-value transfers, with roughly €108 million involved and around €36 million reportedly unrecovered. The details remain a matter of reporting, but the control lesson is clear. A familiar voice, a convincing profile photograph and a plausible message can reinforce one another without proving that the request is genuine. The reported Fideuram case is therefore best understood as a workflow failure, not simply a failure to recognise a deepfake.
That distinction matters. The employee may have behaved exactly as the workflow encouraged: respond quickly to a senior person, continue in the channel already being used, and treat confidence as evidence. If the final check is “does this sound like the executive?”, the attacker only needs to make the voice familiar enough. The stronger question is: “Has this payment instruction been independently confirmed through a channel that the requester did not control?”
Test the decision, not the audio
For a beneficiary change, urgent payment or VIP servicing request, an effective control should require the employee to pause the transaction and contact the purported requester through a trusted route. That might be a known telephone number held in the firm’s directory, a separately initiated call to the client, or a second-person approval using an established procedure. Replying to the same WhatsApp account, calling the number supplied in the message, or asking questions inside the same voice session does not create independence.
This is also why awareness training alone is insufficient. Employees can be told that cloned voices exist, but the organisation still needs to know whether they will challenge a credible request under pressure, record the escalation, and stop the workflow when the independent check cannot be completed. Guidance for the public on AI voice cloning scams reinforces the practical principle: familiarity should trigger verification, not authorisation.
The test should be authorised, controlled and evidence-led. A red team can place realistic calls to staff involved in payment approval, beneficiary maintenance, password resets and executive support. The important outputs are not a headline detection rate. They are the people and steps that held, the point at which the workflow broke, whether escalation occurred, and whether the evidence can stand up to review by the board, an insurer or a regulator.
The board question
The threat is not limited to one reported case. Deepfake voice-fraud attempts rose 1,300 percent in 2024, per Pindrop’s 2025 Voice Intelligence & Security Report, although that is single-vendor telemetry. FinCEN Alert FIN-2024-Alert004, issued on 13 November 2024, also warns that synthetic responses can defeat live prompts, making verification alone insufficient.
The right question for a regulated firm is therefore simple: if a trusted voice requests a payment today, can our people prove the instruction independently before money moves? The DeepBlocker Vishing Readiness Assessment answers that with a fixed-scope authorised engagement and a board-ready evidence pack.
Subscribe to the intelligence for practical tests of whether your voice and payment workflows would survive an AI-enabled fraud attempt.