AI Fraud in 2026: The Honest State of Play

In most technological arms races, the attackers move faster than the defenders. The defenders have institutional processes, compliance requirements, and legacy infrastructure to work around. The attackers have nothing to lose and everything to gain from the next exploit.

AI has accelerated both sides of this equation. Which is uncomfortable, but also somewhat reassuring compared to the alternative: AI accelerating only one side.

The era of bulk credential stuffing and mass phishing (spray paint everything, some of it sticks) hasn't ended, but it has been joined by a different threat model. Fewer attacks, more targeted, more convincing, and significantly harder to catch using the pattern-matching approaches that worked three years ago.

The Synthetic Identity Problem

Synthetic identity fraud is now estimated to cost businesses $20–40 billion globally per year (roughly £16–32 billion, depending on which end of the range you believe). The phrase covers a range of approaches, but the 2026 version is considerably more elaborate than earlier iterations.

The baseline approach of combining real and fabricated identity elements to create a "Frankenstein identity" that bypasses KYC checks has been around for years. What's new is the depth of legend that AI enables. In 2026, a sophisticated synthetic identity doesn't just have a valid-looking credit history. It has a social media presence with AI-generated photos and a years-long post history. A professional LinkedIn profile. In some cases, a small website. The fabricated person exists in enough digital contexts that the identity passes human review, not just automated checks.

The UK Government's estimate that 8 million deepfakes were shared in 2025, up from 500,000 in 2023 reflects the same underlying shift. A 1,500% increase in two years; the kind of figure that looks like a typo but isn't. Meanwhile, UK Finance's Annual Fraud Report put total UK fraud losses at £1.17 billion in 2024. Flat on the previous year, which sounds like good news until you note that banks simultaneously prevented a further £1.45 billion of attempted fraud, up 16% on 2023. The volume of attempts is increasing; the losses are being held roughly steady by active countermeasures.

The cost of creating a convincing fake identity has collapsed. What previously required significant technical skill and resources is now available as a service. Fraud-as-a-Service platforms are accessible for as little as $50 per month (roughly £40), bundling synthetic ID kits, AI-generated voice tools for social engineering, and phishing infrastructure into subscription tiers that resemble nothing so much as a SaaS pricing page.

For ecommerce specifically, this matters most at the account creation and payment onboarding stage. Retailers and payment processors who rely on document verification and face matching to establish account legitimacy are facing adversaries who can inject AI-generated liveness checks that defeat those verification methods. The deepfake injection attack on liveness verification, where the camera stream a verification system sees is replaced in real time with a generated video, is no longer a theoretical vulnerability. It has been documented in production environments, including a Dutch bank case in which a deepfake bypass opened 46 accounts before detection.

The Agentic Fraud Problem

There's a newer concern that emerged in parallel with the agentic commerce rollout: coordinated fleets of AI agents conducting high-speed, multi-step fraud attacks. The same agentic infrastructure that enables a legitimate consumer to delegate purchases to an AI can be used to run automated account-takeover attacks, loyalty point harvesting, or returns fraud at a scale and speed that overwhelms traditional detection systems.

Loyalty point fraud is a specific concern for UK retailers. The major UK loyalty programmes (Nectar, Tesco Clubcard, Boots Advantage) hold significant monetary value, and automated attacks that accumulate and liquidate loyalty currencies are a growing problem. The human fraud analyst reviewing suspicious account behaviour is at a systematic disadvantage against an agent that can probe, test, and exploit at machine speed.

Worth noting that this isn't a failure of agentic commerce as a concept. It's an argument for making sure your fraud and risk teams are involved in the agentic deployment conversation from the start, rather than discovering the new attack surface after the fact. Experian flagged this in its January 2026 fraud forecast as one of its top five threats for the year. The forecast tends toward the dramatic in its framing, but on this particular point the underlying concern is real.

Where the Defences Are Holding

The fraud prevention industry has been deploying AI at scale for several years, and the defensive tools have matured significantly. The approaches that appear to be working in 2026:

Behavioural biometrics (typing rhythm, mouse movement, scroll patterns, device handling) captures signals that are very difficult to fake at the application level. A synthetic identity with a convincing document and face match can still be distinguished from a legitimate user through interaction behaviour, because the patterns of a fraud bot navigating a site don't match the patterns of a human being doing the same thing. The bot is scripted; the human is hesitant, inconsistent, and occasionally distracted. These differences are detectable.

Network analysis and device fingerprinting that look beyond individual sessions to the broader pattern of associated activity. Fraud operations tend to reuse infrastructure: IP ranges, device configurations, account characteristics. That reuse becomes visible when you're looking at population-level patterns rather than individual sessions.

Velocity and graph-based fraud detection that looks at the relationships between accounts, devices, and payment methods across the full customer population, not just at individual transactions in isolation. An account created two weeks ago, with only one prior transaction, connected through a device fingerprint to forty other accounts that all did one transaction at the same retailer in the same week: that's a pattern that rule-based systems miss and graph-based ML does not.

The Honest State of Play

I don't think the fraud picture is catastrophically bad. The tools exist, the investment is there, and the detection rates for most common attack types are meaningfully better than they were three years ago. But the threat environment is also genuinely more sophisticated than it was three years ago, and the asymmetry between cost-to-attack and cost-to-defend has not resolved in the defender's favour.

The retailers most at risk are probably those whose fraud prevention investment hasn't kept pace with their AI commerce investment. The agentic checkout rollout, in particular, opens new attack surfaces that need assessment. What does it mean for your fraud risk when purchases can be made without human verification of intent? What are the implications for your dispute and chargeback processes when the nominal purchaser is an AI agent? These aren't rhetorical questions. They're due diligence items.

Cifas's Fraudscape 2026 report recorded 444,993 cases filed to the UK's National Fraud Database in 2025, a record and a 6% increase on the previous year. Identity fraud accounted for 54% of those filings. The volume is up; the sophistication is up; the tooling available to attackers has never been cheaper or more capable.

None of which is a reason not to deploy agentic commerce, or to avoid AI-assisted personalisation, or to retreat from any of the capabilities that are genuinely improving retail outcomes. The fraud risk is manageable. The industry has the tools. What it doesn't always have is the organisational will to make fraud prevention a first-class concern in the AI deployment conversation, rather than something the risk team finds out about afterwards.

The fraud landscape in 2026 is, to borrow a phrase from a classic British sci-fi series, not fine, but manageable. The "probably" from my February draft I'll leave out. The tools are there. The question is whether the process is.

---

Data sources: UK Finance Annual Fraud Report 2025 (covering 2024 data); Cifas Fraudscape 2026; UK Government deepfakes announcement, February 2026; Persona synthetic identity analysis; Sumsub fraud trends 2026; Experian 2026 Fraud Forecast. For a deeper look at the detection technology side, see AI vs Fraud: The Arms Race in Detection.