A familiar voice can make an unusual request feel routine. A supplier asks for a bank-detail change. A senior executive calls about an urgent payment. A colleague wants a password reset completed before a meeting. The pressure is not necessarily technical. It is social, procedural and time-sensitive.
Recent practical guidance on voice cloning makes the same point: an AI-generated voice can be used to make an urgent request appear to come from a trusted person, particularly where the recipient already recognises the caller. See the guidance from Intellect IT and Vanausdall.
For regulated firms, the question is therefore not simply whether an employee can identify a synthetic voice. It is whether the payment or servicing workflow remains safe when the caller sounds familiar, knows the right names and creates credible urgency.
Training starts the process, but does not prove the control
Awareness training is useful. It establishes the expected response and gives staff permission to pause. But a completion record does not show that a person will challenge a convincing caller when a high-value transaction is waiting, a relationship feels important or a senior colleague appears impatient.
The Arup incident illustrates the distinction. In early 2024, a finance employee wired about 25 million US dollars after a video call involving AI-deepfaked recreations of the company CFO and colleagues. Arup CIO Rob Greig described the event as technology-enhanced social engineering. No systems were compromised and no data was affected, according to CNN Business, the World Economic Forum and Deloitte, as cited in the FinCEN alert. The initial contact was a phishing email, followed by the deepfake call.
FinCEN Alert FIN-2024-Alert004, issued on 13 November 2024, endorses live audio and video verification but warns that attackers can generate synthetic responses to live prompts. It is advisory guidance, not a binding rule. Its practical message is important: verification alone can be defeated, so the surrounding workflow must resist impersonation as well.
Test the behaviour, then preserve the evidence
A resilient control set should be explicit and rehearsed:
- Call back using a trusted number already held in the firm’s records, not a number supplied during the call.
- Require dual approval for payment release and beneficiary changes, with separation between request, verification and execution.
- Use an out-of-band confirmation for unusual instructions, including a second channel and a known contact.
- Define when staff must stop, escalate and record the event, without allowing seniority or urgency to bypass the process.
- Review whether the control worked, not merely whether the employee remembered the policy.
The evidence should show the scenario, the people and workflows tested, the decisions made, the control failures, the remediation owners and the retest position. That is more useful to a board, insurer or regulator than a training attendance report.
DORA, live since 17 January 2025, requires threat-led penetration testing under Article 26 to address people, processes and technologies, with ethical boundaries for social engineering. The UK Payment Systems Regulator’s APP reimbursement regime, live since 7 October 2024, requires reimbursement of eligible victims up to £85,000 per Faster Payments scam claim. These developments make tested resilience a governance matter, not simply a training matter.
DeepBlocker’s Vishing Readiness Assessment is a fixed-scope, authorised two to three week engagement. We place controlled calls against defined workflows and return a sealed, board-ready evidence pack showing where the process held and where it broke.
Take the free demo to see what evidence your current controls would produce.