Deepfake-Enabled Social Engineering: What Security Teams Should Prepare For

What has been observed

A finance worker and a video call. Arup, the London-based design and engineering firm, confirmed to CNN that fake voices and images were used in a fraud against its Hong Kong office. According to Hong Kong public broadcaster RTHK, citing police, a finance worker sent HK$200 million, about US$25.6 million, across 15 transactions. The worker first suspected a phishing email asking for a secret transaction. The doubt ended after a video call with people who appeared to be the chief financial officer and other colleagues. Police said these were deepfake recreations. Arup said its internal systems were not compromised and its operations were unaffected.

The detail that matters is the order of events. The worker’s suspicion was correct. The call was used to remove it.

Official voices. The FBI warned of an ongoing campaign in which text messages and AI-generated voice messages impersonated senior US officials. The aim was to build rapport with targets, then move them to another messaging platform and send a link that harvested login credentials. Compromised accounts were then used to approach the official’s contacts.

Live video and cloned voices. A separate FBI public service announcement on generative AI fraud lists AI-generated video used in live calls to pose as company executives, law enforcement or other authority figures, and cloned voices used to impersonate people to reach bank accounts.

What changes, and what does not

The attack chain stays familiar: pretext, urgency, request, transfer or credential. Deepfakes act on one point in it. The request that once looked suspicious now arrives with a face and voice that match, so the victim’s own judgment, the control the attacker needs to defeat, is supplied with false evidence. The Arup case is a payment fraud with a new tool at the trust step, not a new class of fraud.

Where it enters the organization

What security teams should prepare

1. List the irreversible actions. Payments, bank-detail changes, credential and MFA resets, privileged access grants and data releases. These are the requests where voice or video alone must never be sufficient.

2. Verify through a channel you control. The FBI’s guidance is to look up a known number for the person and call back independently. The same logic applies to anything the requester supplies, including a meeting link.

3. Remove “urgent and confidential” as an accepted exception. The pretext that kept the Arup worker quiet is the one a process should refuse to honor. Staff need a route to escalate a secret request without penalty.

4. Test the callback under pressure. Run authorized tabletop and simulation exercises in which the “executive” pushes back when asked to verify. The failure to look for is people dropping the process when someone senior sounds annoyed. Agree consent, legal review and HR handling before any live exercise.

5. Treat detection tools as a supplement. Vendor claims about deepfake detection have not been independently verified here. Test any product against your own channels before relying on it, and keep the process controls regardless.

6. Plan the response before it happens. If a transfer has gone, treat it as payment fraud: contact the bank for recall at once and report to law enforcement. In the US the FBI asks for reports to IC3.

What is probably not worth building on

Limiting public audio and video of executives reduces raw material for cloning, and it is reasonable hygiene, but it is hard to achieve and gives limited protection. The FBI suggests limiting public sharing of images and voice in its advice to individuals. A program for an organization should not rest on it.

What is not known

Public evidence centers on a handful of reported cases, and organizations rarely disclose attempts that failed. We have not found an authoritative measure of how often deepfake-enabled attempts occur or succeed, so this brief makes no prevalence claim. The reporting on the Arup case does not disclose which tools the attackers used.

Next step

Pick one process from the table, such as changing a supplier’s bank details, and walk it end to end asking one question at each step: is there anything here an attacker could satisfy with a convincing voice or face? Phishing Operations exercises can then test whether the control holds when a believable executive pushes back.

Sources

1. CNN. British engineering giant Arup revealed as $25 million deepfake scam victim. edition.cnn.com

2. FBI Internet Crime Complaint Center. Public service announcement on senior US officials impersonated in a malicious messaging campaign. ic3.gov

3. FBI Internet Crime Complaint Center. Public service announcement: criminals use generative artificial intelligence to facilitate financial fraud. ic3.gov