Deepfakes and CEO fraud: generative AI revives social engineering

Up to date as of October 31, 2026. The AI Act transparency obligations have applied since August 2, 2026; the transition period for marking generated content ends on December 2, 2026.

In January 2024, an employee at the Hong Kong office of the engineering group Arup joined a video call with the group’s chief financial officer and several colleagues. Following their instructions, they made fifteen transfers, totaling 200 million Hong Kong dollars (about 25 million US dollars). Every other participant in the meeting was a deepfake, built from public videos and recordings. Arup confirmed the incident in May 2024.

CEO fraud is not new. What has changed is that generative AI makes it much more convincing, and much cheaper to set up.

What AI changes

Credibility: a few public clips (interviews, conferences, corporate videos) are enough to imitate an executive’s voice. Faces can be animated live in a video call. Messages are written without mistakes, in any language, in the tone of the person being imitated.
Scale: fraudsters use public information (organization charts, professional profiles, company news) to personalize their attacks at low cost.
Multiple channels: an email is confirmed by a phone call, then by a video call or an instant message. Each channel reinforces the credibility of the others.

The most common scenarios

CEO fraud: a fake executive asks for an urgent and confidential transfer, often for an alleged acquisition or a sensitive deal.
The fake supplier: a usual partner announces a change of bank details.
The fake employee calling the IT help desk: the fraudster impersonates an employee to get a password or a multi-factor authentication reset, then gets into the information system.

The warning signs are the same as before AI: urgency, confidentiality, bypassing procedures, an unusual channel, new bank details and refusal of a call-back.

Why technology will not be enough

What the AI Act provides

Article 50 of the AI Act imposes two useful obligations:

providers of generative AI systems must mark their content in a machine-readable and detectable way;
professionals who publish deep fakes must disclose that the content has been generated or manipulated.

These obligations have applied since August 2, 2026. Systems placed on the market before that date have until December 2, 2026 to comply with the marking obligation. To help organizations comply, the European Commission published a voluntary code of practice in June 2026.

In France, the SREN law of May 21, 2024 (on securing and regulating the digital space) also extended the offense of making montages without consent to deep fakes generated by algorithms.

The limits

These rules govern legitimate uses. A fraudster, on the other hand, will put no label on a fake video call, and can use misused or non-compliant tools. Deepfake detection tools are improving, but they remain imperfect and locked in a permanent race with generation tools.

Protection therefore rests first and foremost on procedures and on the reflexes of the teams.

Procedures that hold up

No transfer on the strength of a single exchange, whatever the channel (email, phone call, video call): approval by two people above a defined threshold.
A systematic call-back on a known number, taken from the internal directory, never the one given in the message.
A verification question or code agreed offline for sensitive requests coming from management.
A strict procedure for changes to suppliers’ bank details: independent verification and a waiting period before the first payment.
Stronger identity checks at the IT help desk, especially for resetting access to privileged accounts.
An explicit right to verify: management states that no employee will ever be blamed for checking a request, even one that comes from management.
Controlled public exposure: information about the organization, internal procedures and executives’ travel.

Raising awareness without spreading fear

Awareness training works when it is concrete: real examples, simulated phishing and fraudulent calls, lessons learned after each detected attempt. The groups to target first are accounting and treasury, executive assistants, the IT help desk and human resources.

If the fraud succeeded

Notify your bank immediately to try to block or recall the funds: every hour counts.
File a complaint, and quickly: to be compensated by its cyber insurance, a French company that is the victim of a cyberattack must file a complaint within 72 hours.
Keep the evidence: messages, recordings, connection logs.
Notify your insurer and get support, for example through Cybermalveillance.gouv.fr, the French government’s assistance platform for victims.

These steps come on top of your other notification obligations, detailed in our analysis One incident, several clocks.

Sources

← All insights