What is a deepfake?
A deepfake is synthetic audio, video, or imagery generated by artificial intelligence to convincingly imitate a real person’s face, voice, or both. The EU AI Act defines it in law as AI-created or AI-altered content that imitates real people or events in a way that could mislead. In a business setting, that usually means a cloned executive voice on a call or a fabricated face in a video meeting.
For years deepfakes were a celebrity and politics problem. That has changed. Criminals now use the same tools to impersonate the people you work with, your CFO, a supplier, a colleague, and to authorise payments that should never happen. The technology is cheap, fast, and built from footage your executives have already posted in public.
How deepfakes are made
Modern deepfakes are produced by generative AI models trained to reproduce a target’s likeness. Voice cloning systems learn a person’s tone and cadence from a short sample. Video models map a target’s face onto a live video feed in real time, so an attacker can sit on a call and appear as someone else as they speak.
The barrier to entry has collapsed. The World Economic Forum reports that cloning a usable voice takes only 20 to 30 seconds of clear audio, and a convincing video deepfake can be built in under an hour using freely available software. The source material is scraped from the places your leaders are most visible: LinkedIn, YouTube, recorded webinars, and earnings calls. After the Arup fraud, the firm’s own technology chief told the Forum he recreated himself as a real-time deepfake out of curiosity, using free open-source tools.
Types of deepfake attack
Deepfakes show up in several distinct attack patterns. Knowing the shape of each one makes it easier to recognise in the moment.
- Voice cloning (vishing): a cloned voice on a phone call or voicemail pressuring someone to pay or share information.
- Live video impersonation: a real-time face swap on a video meeting, impersonating an executive or several colleagues at once.
- Lip-synced or puppet video: existing footage altered so a real person appears to say something they never said.
- Synthetic identity and document fraud: generated faces and ID images used to defeat facial recognition and pass identity checks.
- Off-camera impersonation: a faked persona in a meeting chat window or messaging app, often combined with a cloned voice elsewhere in the attack.
The business impact
Deepfake fraud has moved from novelty to measurable loss. Reported losses from deepfake-enabled fraud reached roughly 410 million US dollars in the first half of 2025 alone, more than the 359 million reported for all of 2024, according to analysis by Surfshark. The Deloitte Center for Financial Services projects that generative-AI-enabled fraud in the United States will climb from 12.3 billion dollars in 2023 to 40 billion by 2027.
The damage is not only the money that leaves the account. A single successful attack carries several costs at once:
- Direct payment fraud: funds wired to attacker-controlled accounts, often unrecoverable.
- Supplier and invoice fraud: changed bank details confirmed by a faked voice, redirecting legitimate payments.
- Identity and access bypass: synthetic faces defeating biometric checks. Entrust reported that around one in five biometric fraud attempts in the past year were deepfakes.
- Reputation and reporting risk: loss of confidence, regulatory exposure, and the cost of investigation and remediation.
Deepfakes also make the oldest fraud category harder to stop. Business email compromise still drove 3.05 billion dollars in reported losses to the FBI in 2024. A cloned voice or face on top of that email removes the moment of doubt that used to protect people.
Real-world cases
Three cases from 2024 show the same playbook and, importantly, the single control that decided each outcome.
Arup, about 25 million US dollars lost. A finance employee at the engineering firm’s Hong Kong office joined a video call with people who looked and sounded like the UK-based CFO and several colleagues. Every participant except the employee was a real-time deepfake, built from public footage. He made 15 transfers totalling around 200 million Hong Kong dollars before he contacted head office and learned the meeting never happened. The control that would have stopped it: verifying the payment out of band, through a known channel, before sending a single transfer, not after.
WPP, attack failed. Fraudsters cloned the voice of CEO Mark Read and used a fake WhatsApp account and a Microsoft Teams meeting to target an agency leader, impersonating Read in the chat window. The attempt failed because the targeted executive recognised the red flags, a secret new venture, requests for money and personal details, and verified instead of acting. The control that worked: trained vigilance. As Read put it to staff, just because an account has his photo does not mean it is him.
Ferrari, attack failed. An attacker cloned the voice of CEO Benedetto Vigna, accent and all, in an attempt to push through an urgent request. An executive ended the call by asking a question only the real Vigna could answer. The control that worked: a personal verification question agreed in advance.
Deepfakes and compliance
Deepfake fraud is now woven into several regulatory regimes. The duties below are the ones a Swedish board most needs to understand.
NIS2 and Cybersäkerhetslagen. Sweden transposed the EU NIS2 Directive into national law as Cybersäkerhetslagen (SFS 2025:1506), in force since 15 January 2026. Article 21 requires security awareness training and incident handling, the measures that address social-engineering attacks like these. Under Article 20 the board is accountable for those measures, and supervisory authorities can hold board members personally responsible. Read our NIS2 compliance guide for Sweden for the full obligation set.
The EU AI Act. From 2 August 2026, Article 50 of the AI Act requires deepfakes to be clearly labelled as artificially generated, and the rule applies even when there was no intent to deceive. Breaching the labelling duty can cost up to 15 million euros or 3 percent of global turnover. This duty targets legitimate deployers of AI, not criminals, so it helps the wider information ecosystem rather than stopping a fraudster directly.
DORA and GDPR. For financial entities, DORA Article 17 governs ICT incident management, supervised by Finansinspektionen. If a deepfake attack leads to a breach of personal data, GDPR Article 33 still requires notification to IMY within 72 hours. See our DORA compliance overview for the financial-sector detail.
How to spot a deepfake
There are tells, but treat them as hints, not proof. Real-time deepfakes increasingly pass visual inspection, and the people fooled at Arup were professionals looking at familiar faces.
- Visual: unnatural or absent blinking, lip movements slightly out of sync, odd lighting at the hairline, blurred or warping edges around the face, a stiff or fixed gaze.
- Audio: flat intonation, unusual pacing or breathing, a slight delay between audio and video, a robotic quality under emotion.
- Behavioural: urgency, secrecy, pressure to bypass normal process, a request routed through an unusual channel, and reluctance to switch to a phone call or do a callback.
The honest caveat is that detection by eye is unreliable and getting worse. Gartner classes deepfake detection tools as necessary but not sufficient on their own, because attackers test against them and new techniques slip through. Treat verification through a separate channel, not your own perception, as the real safeguard.
How to defend against deepfakes
Deepfake fraud is defeated by process and habit, not by spotting the fake. Put the controls below in place across the teams that move money and grant access.
People. Train finance, HR, and payment-approval staff on this specific attack pattern, and make it safe to pause and verify even when the caller is senior and the matter is urgent. Brief executives that their public footage is the raw material for these attacks.
Process. This is where attacks are stopped:
- Require dual authorisation for payments above a set threshold.
- Verify unusual or high-value requests out of band, through a pre-agreed channel such as a callback to a number you already hold. Never verify through the channel the request arrived on.
- Agree a personal verification question or code word for sensitive requests.
- Control supplier bank-detail changes by confirming on known contact details, never details supplied in the request.
- Adopt a simple rule that no transaction is ever too secret or too urgent for normal verification.
Technology. Add liveness and anti-spoofing to biometric checks, deepfake detection at the contact centre and in identity verification, and email authentication such as DMARC to cut the spear-phishing that usually starts the chain. Remember Gartner’s point: these tools support the process, they do not replace it. If you only do one thing, make every high-value or unusual request verified through a second, known channel before anyone acts.
This post is also available in:
Svenska
