Cheaper lies, faster scale: how AI is used for fraud in 2026
Artificial intelligence has not invented a new kind of greed. It has made old lies cheaper to produce and easier to scale. A phishing email can now be rewritten in fluent, official-sounding English in seconds. A voice can be cloned from a short public clip. A video call can be populated with faces that look like colleagues. None of that guarantees that a particular message is a scam. It does mean that “it sounded like them” is a weaker test than it used to be.
The useful way to read 2026 is not as a year in which every fraud became a deepfake. Most losses still come from familiar typologies: investment scams, authorised payments, romance and confidence tricks, business email compromise. What has changed is the industrial layer underneath them. Generative tools have made social engineering more personal, more multilingual and more abundant. The artefacts are getting harder to spot by instinct alone. The defences that still work are mostly boring: a callback on a number you already have, a second person to approve a payment, and a habit of treating urgency and secrecy as warning signs rather than as proof that the request is important.
What actually changed
Europol’s Internet Organised Crime Threat Assessment for 2026, published in April, is direct about the direction of travel. Online fraud schemes, it says, remain the fastest-growing area of organised crime, commonly involving phishing and an infrastructure built for speed and industrial-scale victimisation. Generative AI has improved fraudsters’ ability to personalise social engineering. Law-enforcement reporting through 2025 pointed to a shift from a landscape defined mainly by volume and variety to one also marked by velocity. Europol says generative AI allowed networks operating from outside the EU to target EU citizens at high speed and vast scale, including in AI-assisted impersonation of bank helpdesk or law-enforcement personnel, and in more realistic business email compromise and CEO fraud.
The same report is cautious, and that caution is worth keeping. While generative AI is already in use, Europol warns that the criminal adoption of agentic AI — systems that can automate parts of the criminal process — is set to raise the threat further. The dark web, it adds, remains a hub for malicious large language models built or adapted to strip out the ethical constraints of commercial tools. The product being sold is not a cinematic deepfake in every case. Often it is a better script, a more convincing email, or a chatbot that can hold a long conversation with thousands of targets at once.
ENISA, the EU’s cybersecurity agency, reached a similar conclusion in its 2026 Threat Landscape, which reviewed incidents observed in 2025. It found an increasing use of AI by malicious groups, primarily to facilitate or enhance their activities, and noted that synthetic audio and video and AI-generated text have become part of the daily arsenal of some information-manipulation actors. AI is, in other words, mostly a force multiplier. It does not have to be perfect. It only has to be good enough, often enough, to get someone to move money before they check.
The video call that should have been a callback
The case that still best illustrates the corporate end of this problem happened in January 2024, and it has not aged into a historical curiosity. In a written reply to Hong Kong’s Legislative Council on 26 June 2024, the Secretary for Security said police had received a fraud report involving a pre-recorded deepfake video conference. The victim had received a phishing email from someone impersonating the chief financial officer of the firm’s United Kingdom head office, and was invited into a group video conference about purported confidential transactions. The employee authorised transfers to five local bank accounts and lost about HK$200 million. Police believed the conference had been generated from publicly available video clips and voices. Because the meeting was pre-recorded, there was in fact no live interaction. After giving the instruction, the fraudster ended the meeting on some pretext and continued the payment instructions over instant messaging.
The engineering firm Arup later confirmed to CNN that it was the victim, that it had notified Hong Kong police in January, and that fake voices and images were used. A company spokesperson also said its financial stability and operations were not affected and that no internal systems were compromised. The lesson is not that every video call is fake. It is that a familiar face on a screen, combined with a demand for secrecy and speed, is no longer a sufficient basis for releasing funds.
Hong Kong police recorded a second case in May 2024, in which a staff member of an unnamed multinational joined a WhatsApp approach and then a video conference lasting nearly 30 minutes, and was instructed to transfer nearly HK$4 million. Police again believed publicly available video had been altered and played into the call.
The same pattern has since been alleged at much larger scale. On 5 November 2026, the US Attorney’s Office for the Southern District of New York announced charges against three men for a scheme that, prosecutors allege, used “super intelligence” between 2024 and 2026 to impersonate the voices and images of chief executives and of lawyers at major law firms, and to trick employees into sending more than $129 million to accounts controlled by the conspiracy. The alleged sequence will be familiar to any payments team: a confidential transaction, a cloned executive, a purported law firm, and a wire to an account that was not what it seemed. The proceeds, prosecutors say, were then converted into cryptocurrency. The charges are accusations, and the defendants are presumed innocent unless and until proven guilty. The alleged method is nonetheless one that finance teams should already be designing controls around: do not let a voice, a face or a demand for secrecy replace an out-of-band check.
The numbers that are public
The first year in which a major official statistical series isolated AI as a descriptor is also the year that showed how uneven the picture still is. In its 2025 annual report, the FBI’s Internet Crime Complaint Center recorded 22,364 complaints with an AI-related descriptor, and adjusted losses of $893,346,472. Those figures sit inside a much larger total: losses reported to IC3 in 2025 surpassed $20 billion. AI-tagged cases are therefore a meaningful slice, not the whole of cyber-enabled fraud, and the Bureau itself notes that many victims may not realise when AI was used.
Where complainants did report an AI nexus, the losses clustered in familiar typologies. Investment scams with a reported AI link surpassed $632 million, against overall investment-scam losses of more than $8 billion. The Bureau says scammers used AI to generate thousands of conversations that appeared different to each prospective victim, and to produce videos and voices of celebrities, chief executives or other trusted figures as fake endorsements. Businesses reported losses of more than $30 million to business email compromise involving AI, including chat-generated emails mimicking officials and voice cloning used to request wires. Confidence and romance scams with a likely AI nexus accounted for more than $19 million, including “distress” or grandparent scams in which a cloned voice mimics a loved one in an emergency, with victims reporting more than $5 million in those cases alone. Employment scams with an AI involvement accounted for almost $13 million. In some of those, the Bureau says, lip movement and other on-camera actions did not line up with the audio, and the apparent goal was often access to a private network rather than an immediate cash theft.
Those are US complaint figures, and they understate global harm. They are still the clearest official breakdown currently available of how AI is being used, and for what.
In the United Kingdom, UK Finance reported that its members saw criminals steal £1.28 billion through payment fraud in 2025, up 4 per cent. Authorised push payment fraud — payments the customer is manipulated into making — rose to £576.4 million, up 19 per cent, across 248,070 cases. Investment fraud was the largest APP loss type, at £221.5 million, up 40 per cent. The report does not isolate an AI-only total, and it would be wrong to invent one. What it does show is the environment in which AI-enabled social engineering is most useful: scams that depend on persuading a person to authorise a payment after contact on a platform or a call that appears to come from a trusted institution.
Ireland’s official series tells a similar story. Irish resident payment service providers recorded €160 million in fraudulent payments in 2024, up 24.5 per cent on the year before. That figure is not an AI statistic. It is the pool of losses into which more convincing impersonation can drain.
Scripts at industrial scale, and faces at the onboarding gate
The most common AI fraud of 2026 is probably not a Hollywood-grade executive on a video wall. It is a more convincing conversation, delivered at a volume no human boiler room could sustain. The FBI’s description of investment and romance typologies is, in substance, the same as the long-con model often called pig-butchering: a relationship is built, trust is rehearsed, and only then comes the request to invest, to move funds, or to help in an emergency. Large language models do not have to invent that script. They only have to localise it, keep it consistent, and run it in parallel.
The same synthetic content is also being pointed at identity checks. In August 2023, Hong Kong police, in an operation named Smashscam, dismantled a syndicate suspected of stealing identities and using an AI face-swapping programme to apply for loans online from finance companies, between September 2022 and July 2023, involving HK$200,000. Nine people were arrested. Officials described it as the first detected case involving AI face-swapping. That is a smaller sum than the later corporate video-conference cases, and a clearer warning for anyone running remote onboarding: a “liveness” check that only watches what appears on camera can be shown a face that does not belong to the person at the keyboard.
There are two practical ways that kind of attack can work. In a presentation attack, a fake image, a replayed video or a face-swap is simply shown to the camera. In an injection attack, the video never goes through the real camera at all: a pre-recorded or generated file is fed into the capture stream. A check that only asks a customer to turn their head or smile may catch some of the first type and miss the second. Compliance teams should assume that both will be tried, and should not treat a passed selfie test as conclusive on its own. Device-path integrity, injection detection and slower step-up checks for higher-risk applications are the unglamorous counterparts to a glossy “AI liveness” label.
How to slow it down
If you are the person being asked to pay or to share access
The strongest control is still out of band. If a caller, a video meeting or a message asks you to transfer money, change payment details, buy crypto, or hand over a one-time passcode, break the channel. Call the person or the firm back on a number you already have, not on a number supplied in the same conversation. Hong Kong police, briefing the public on deepfake deception, have made the same points in official advice: ask the person on camera to make movements and watch for abnormalities; test them with conversation that a recording cannot handle; if a relative or friend asks for a remittance by video or audio, verify it with a phone call; and treat unexpected video calls with caution.
UK Finance’s “Take Five” advice is as good a short checklist as any: stop, challenge, and only then protect. Criminals, it notes, are practised at impersonating people, organisations and the police, and will try to rush or panic you. Only a criminal needs you to decide before you can check. If money has already left the account, contact your bank immediately and report the matter. In Ireland, that means your bank and An Garda Síochána. In the United Kingdom, UK Finance points people to Report Fraud. The Central Bank of Ireland has found that reporting still changes outcomes. In a consumer study cited in its July 2026 Financial Crime Bulletin, 38 per cent of fraud victims never reported the experience to their provider or to any authority. Among those who did report to their bank, Gardaí or another authority, 57 per cent recovered money, compared with 13 per cent of those who did not.
If you work in fraud, KYC or compliance
Design the process so that a familiar voice cannot finish the job. Payments above an agreed threshold should need dual control. Callbacks should be made to a pre-registered number, not to a number in the email or on the call. Voice and video should not be treated as authentication. Staff should be told, explicitly, that a request for secrecy is a reason to escalate, not a reason to bypass a control. For remote onboarding, “liveness” is not enough if the capture path can be injected; test for both presentation and injection attacks, and add friction when a new account, a new device or a new payee appears in combination. Behavioural signals still matter: unusual urgency, coaching in the background, a customer who cannot deviate from a script, or a payment path that heads straight for a mule account or a crypto off-ramp.
AI has made the performance more convincing. It has not repealed the oldest rule in payment fraud. If the request is urgent, secret, and requires you to move money or access now, it is exactly the moment to slow down.
Further watching
The Hong Kong video-conference case, and the later allegations in New York, are easier to assess once you have seen how cheaply a convincing face and voice can now be made. In "The rise of deepfakes and how to stop them", a Financial Times film, AI correspondent Melissa Heikkilä walks through how deepfakes are built, dissects scams that used them — including one that targeted the FT's own Martin Wolf — and looks at what detection, and ordinary procedure, can still do. The practical conclusion sits close to the one in this article: a familiar face or voice on a call is no longer, on its own, a reason to move money.
Watch "The rise of deepfakes and how to stop them | FT Film" by Financial Times on YouTube
Sources
- Europol, The evolving threat landscape. How encryption, proxies and AI are expanding cybercrime – Internet Organised Crime Threat Assessment (IOCTA) 2026: https://www.europol.europa.eu/cms/sites/default/files/documents/IOCTA-2026.pdf, and the IOCTA landing page: https://www.europol.europa.eu/publications-events/main-reports/iocta-report.
- ENISA, “Exploring the evolution of the cyber threat landscape: How dependencies weaken our digital resilience” (22 September 2026), presenting findings from the ENISA Threat Landscape 2026: https://www.enisa.europa.eu/news/exploring-the-evolution-of-the-cyber-threat-landscape-how-dependencies-weaken-our-digital-resilience.
- Hong Kong Government, Legislative Council written reply, “LCQ9: Combating frauds involving deepfake” (26 June 2024): https://www.info.gov.hk/gia/general/202406/26/P2024062600192p.htm.
- CNN Business, “Arup revealed as victim of $25 million deepfake scam involving Hong Kong employee” (16 May 2024), including Arup’s confirmation that fake voices and images were used: https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk.
- US Attorney’s Office, Southern District of New York, “Three Foreign Nationals Arrested And Charged With Scheme To Use Super Intelligence To Defraud Companies Of $129 Million” (5 November 2026): https://www.justice.gov/usao-sdny/pr/three-foreign-nationals-arrested-and-charged-scheme-use-super-intelligence-defraud. The charges are accusations, and the defendants are presumed innocent unless and until proven guilty.
- Federal Bureau of Investigation, Internet Crime Complaint Center, 2025 Internet Crime Report: https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf.
- UK Finance, “Fraud remains a national security threat as criminals steal almost £1.3 billion” (15 June 2026), accompanying the Annual Fraud Report 2026: https://www.ukfinance.org.uk/news-and-insight/press-release/fraud-report-2026-press-release; report landing page: https://www.ukfinance.org.uk/policy-and-guidance/reports/annual-fraud-report-2026.
- Central Bank of Ireland, “Payment Fraud Statistics” (statistical release, 16 October 2025): https://www.centralbank.ie/docs/default-source/statistics/data-and-analysis/payment-fraud-statistics/payment-fraud-statistics-202485f0d58e-4a6d-4a27-aed4-03717c7bdcd7.pdf?sfvrsn=3a96e1a_1; and Financial Crime Bulletin, Issue 2 (July 2026): https://www.centralbank.ie/docs/default-source/publications/financial-crime-bulletin/financial-crime-bulletin---issue-2-july-2026.pdf.