Deepfake fraud starts with data: what attackers collect first

What is a deepfake CEO scam?

A deepfake CEO scam is a fraud in which criminals use an AI-generated voice or video of a senior executive to make an employee send money, share a login code or open access. It is the synthetic-media version of CEO fraud and business email compromise.

The fake is the last step. Before it, the attacker needs material and context: what the executive sounds like, who approves payments, which phone numbers and email addresses are in use, and what the executive is doing this week. That input is data, and much of it already sits in public sources and in breach and leak data. This article follows that data trail. For which exposed fields make a plain email scam convincing, see CEO fraud and BEC: which exposed data fields make it convincing. For a test of how far an AI-written lure gets, see the AI CEO impersonation phishing red team experiment.

What did the Hong Kong deepfake case need to work?

The Hong Kong case of January 2024 needed public recordings of a real executive, a believable story and a finance employee who could authorize transfers. Hong Kong police say the fraud cost about HK$200 million, and that the fake video meeting was built from downloaded public video clips and the voice of the impersonated officer.

The details come from a written reply by the Secretary for Security to the Legislative Council on 26 June 2024. In end-January 2024, the Police received a fraud case in which the informant got a phishing email from someone posing as the Chief Financial Officer of the head office in the United Kingdom and was invited to a group video conference about confidential transactions. As instructed, the informant authorized transfers to five local bank accounts. The Police believed the video conference was pre-recorded and generated "using downloaded online public video clips and voices of the impersonated officer", so no real interaction took place.

The company was later identified as the engineering firm Arup. In May 2024 a spokesperson told CNN that Arup had notified Hong Kong police in January and could "confirm that fake voices and images were used". Its global chief information officer, Rob Greig, said attacks including "invoice fraud, phishing scams, WhatsApp voice spoofing, and deepfakes" were rising in number and sophistication.

Two points follow for defenders. First, the reply describes a pre-recorded meeting, so the employee could not have tested it with live questions. Second, nothing in the description needs a breach: public clips and an org chart were enough to start.

What do attackers collect before they build the fake?

Attackers collect voice and video samples, the names and roles around the executive, the phone numbers and email addresses in use, the writing style, and the calendar. Each item makes the later call or video meeting easier to believe, and each has a common source and a way to reduce it.

The table below is our synthesis. The Hong Kong reply and the FBI notices cited here name some of these inputs directly (public clips, voices, contact information from compromised accounts). The sources column describes where such data usually comes from in general.

The reduction column does not make an executive unrecordable. A keynote stays public. The aim is to remove what is cheap to remove, such as phone numbers, finance contacts and old credentials, so the fake has less to work with. Fewer facts also make the attacker's pretexting easier to catch.

How does breach data feed a deepfake scam?

Breach data feeds a deepfake scam by supplying working contact details and sometimes access. A leaked email address and phone number tell the attacker where to reach the executive and the assistant, and a compromised mailbox or messaging account can reveal who talks to whom about payments.

The FBI describes this loop in a 2025 notice on impersonated senior US officials. Since April 2025, the FBI says, malicious actors have sent text messages and AI-generated voice messages that claim to come from a senior US official. Access to personal or official accounts could be used to target other officials or their contacts "by using trusted contact information they obtain", and contact information obtained through social engineering "could also be used to impersonate contacts to elicit information or funds". A December 2025 update says the activity dates back to 2023 and that victims were asked to provide an authentication code, supply personal documents, or wire funds overseas. The campaign targets government officials, but the pattern applies to company leaders: stolen contacts make a synthetic voice reach the right person.

An exposed password gives the attacker more than a login. It opens the mailbox where the executive's real tone, contacts and approval routine are visible. That is why the first defensive step is knowing which leadership accounts are already in circulation. CEM starts from your company name, finds your executives and checks their current and past email addresses against breach corpora and leak databases, ranked by what to fix first. It answers whether a leadership address or password is already out there. For the reuse pattern behind many of these cases, see password reuse and what to do when an executive's password is in a breach.

Data that was stolen and then posted or sold sits in different places. EDG3 is WebThreat's threat intelligence for that: it covers dark web forums and marketplaces, Telegram and other channels, and leak databases, and shows what has been posted about a domain you have authorized. For how this exposure builds up around one person, see executive digital exposure: the part that sits in breach and leak data and executive personal accounts: what breach data shows.

Can finance teams spot a deepfake by looking and listening?

Finance teams cannot rely on spotting a deepfake by sight or sound. The FBI says AI-generated content "has advanced to the point that it is often difficult to identify", and that a cloned voice and a real voice can sound nearly identical. The reliable control is an out-of-band check that does not depend on how the call looks.

FinCEN, the US Treasury bureau for financial crimes, reached a similar conclusion for banks in its alert FIN-2024-Alert004 of 13 November 2024. It says generative AI tools "have greatly reduced the resources required to produce high-quality synthetic content". FinCEN wrote the alert for financial institutions and customer identity checks, not for executive scams, so we use it only for that point about effort and cost.

Visual tells still exist. The FBI lists distorted hands, irregular faces, inaccurate shadows, voice call lag and unnatural movements, and Hong Kong police suggest asking a caller to turn the head or testing the conversation. Treat these as extra signals. A calm voice and a clean image prove nothing, and a pre-recorded meeting never reacts to a test at all.

Which verification protocol should a finance team use?

A finance team should treat every payment or access request that arrives by voice, video, chat or email as unverified until it is confirmed on a channel the attacker did not supply. The protocol below uses controls the FBI and Hong Kong police recommend, and adds a process control, dual approval, that this article recommends as general practice.

1. Call back on a number you already hold. Do not use the number in the message, the invitation or the chat. The FBI advises independently identifying a phone number for the person and calling to verify.
2. Use a code word for high-risk requests. The FBI recommends a secret word or phrase agreed in advance for family members. Companies can apply the same idea to payment approvals, agreed face to face and not stored in email or chat.
3. Require two approvers for any new payee, changed bank details or urgent transfer. Both approvers confirm separately. A meeting with "the CFO" counts as one request, not as approval.
4. Move verification off the channel of the request. A video call is verified by a phone call to a known number. A text is verified by a call or an internal ticket.
5. Never share codes. The FBI says never to give a two-factor code to anyone over email, text or an encrypted messaging application.
6. Slow down urgent or secret requests. Confidential deal, last-minute transfer and "do not tell the team" are the usual wrapper. Make an escalation path that lets staff stop without being penalized.
7. Treat a move to a new app as a warning. The FBI notes the actors ask to move the conversation to a different messaging platform almost immediately.
8. Report and record. Keep the message, number and meeting link, and report to the bank, internal security and the national reporting channel (in the US, the FBI's Internet Crime Complaint Center).

Brief assistants and executives too. They hold the numbers, calendars and approvals that the attacker is looking for.

Which data should leadership reduce first?

Leadership should reduce phone numbers, exposed credentials and finance contacts first, because those are the inputs that are cheap to change and expensive for the attacker to replace. Voice and video samples come second; most executives cannot withdraw a conference keynote.

The FBI's December 2024 notice on generative AI fraud suggests, where possible, to "limit online content of your image or voice, make social media accounts private, and limit followers to people you know". For an executive this is a sensible trade-off, not a rule: keep official channels public, tighten personal ones.

Personal data that feeds pretext calls, such as home address, family names and mobile numbers, often sits in people-search listings. WebThreat's information removal service covers this: our specialists audit where an executive's personal data appears, file opt-outs and takedown requests, verify each removal and re-check closed listings every 30 days. Families are covered on request. It reduces what a caller can quote back, so a pretext call has less to work with. See also data removal vs breach monitoring: know which one you are buying.

Breach data cannot be deleted from copies already in circulation. The response is to make it useless: new passwords, sign-in methods that cannot be replayed by a caller, and the payment protocol above.

How WebThreat can help

WebThreat does not run your payment approvals or detect fake audio and video. The protocol above belongs to your finance and security teams. Our part is the data that feeds the fake.

• CEM finds your executives from your company name and checks their current and past email addresses against breach corpora and leak databases, ranked by what to fix first. See what CEM finds for your leadership team.
• EDG3 is threat intelligence for stolen data posted and sold on dark web forums, marketplaces, Telegram and other channels, and in leak databases.
• Information removal specialists handle executives' personal data in people-search listings. Ask our specialists about removal.

Questions about which service fits? Contact us or read the methodology.

This article reflects the sources linked above as of 29 October 2026. Case details can change as investigations continue, and this is general information, not legal advice.