Stripe Radar in 2026: Foundation Models, Card Testing Down 64%, and the Adaptive Rules Era

The Foundation Model Shift: Tens of Billions of Transactions

For a decade, fraud detection has been a gradient-boosted tree problem. Vendors stacked XGBoost ensembles, handcrafted features (BIN range, device fingerprint, velocity over 24 hours) and tuned thresholds. The models were small, specialised and brittle — and they stopped improving once the feature store stopped getting richer.

Stripe's 2026 announcement reframes the entire space. Instead of a fraud-specific classifier, Stripe trained a single foundation model for payments on tens of billions of transactions across its network. The same base model now powers fraud detection, authorisation optimisation, dispute prediction and the new Payments Intelligence Suite. Think of it as the payments equivalent of a large language model: one general-purpose learner, many downstream tasks.

That's not marketing gloss. The practical consequence is that every new behaviour Stripe observes — a novel card-testing pattern in Sao Paulo, a fresh synthetic-identity ring on a US platform — updates a shared representation that lifts performance on every downstream task simultaneously. Conventional fraud ML simply can't do that without rebuilding each model.

Card Testing Down 64%: A Specific, Measurable Win

The most concrete win from the foundation model so far is a 64% reduction in card testing attacks compared to Stripe's previous systems. Card testing — where fraudsters hammer checkout pages with stolen card numbers in rapid succession to validate which ones still work — has been one of the fastest-growing fraud categories of the decade. It's automated, cheap, and relentlessly adaptive.

A 64% drop isn't a rounding error. For a mid-sized SaaS business, card testing isn't just a fraud expense — it inflates gateway costs, damages sender reputation with issuers, and can push authorisation rates down across legitimate customers. Stripe's newer model catches the behavioural signature of automated testing earlier in the funnel, before the BIN attack ladder does real damage.

Adaptive Rules: A 1.3pp Lift in Payment Success

The second release is quieter but arguably more important commercially. Adaptive Rules combine Stripe's ML models with live issuer responses — CVC match, AVS postal code response, and more — in real time, per transaction. The outcome published by Stripe: businesses see a 1.3 percentage point increase in payment success rates with minimal fraud impact.

If you don't work in payments, 1.3 points sounds small. It is not. For a business doing £10m in annual online revenue, a 1.3pp auth-rate lift is roughly £130,000 in recovered revenue per year that previously bounced at the issuer. And because Adaptive Rules lean on live issuer signals rather than static thresholds, the uplift compounds as issuers get better at responding.

Radar Assistant: Natural Language to Fraud Rules

Radar has always had a rules engine. The problem: writing rules required SQL-adjacent expertise and a deep understanding of Stripe's event schema. Analysts and ops leads were permanently dependent on engineering.

Radar Assistant closes that gap. Write a rule in English — 'block transactions above $500 from new customers using a prepaid card' — and Radar Assistant translates it into a live, versioned rule that runs inside the Radar runtime. The implications are cultural more than technical: fraud strategy is no longer bottlenecked by engineering availability. A fraud analyst can experiment, deploy, and roll back in minutes.

This is the same pattern we're seeing across every enterprise software category in 2026. The people with the domain knowledge are finally allowed to touch the system that uses it.

Radar for Platforms: Account-Level Fraud Detection

Most fraud discussion focuses on transactions. But for marketplaces, SaaS platforms and anyone running Stripe Connect, the bigger exposure is often account fraud — fraudulent connected accounts set up to launder stolen funds, run bust-out schemes, or abuse payouts.

Radar for Platforms extends Stripe's fraud stack upward: account fraud detection, fraudulent connected account identification, account-level rules, and advanced platform analytics. Crucially, Stripe says the underlying models are trained on data from 14,000+ platforms, which is the kind of cross-platform signal no individual marketplace could ever replicate in-house.

For any business running a Connect-based model — a creator economy app, a B2B marketplace, a vertical SaaS with embedded payments — this is the component that finally makes risk proportional to scale rather than ahead of it.

Stripe vs Adyen vs Checkout.com: The Framing Gap

It's worth being fair here. Adyen, Checkout.com and Braintree all ship sophisticated ML-based fraud detection. Adyen's RevenueProtect, Checkout.com's Risk.js and Braintree's Fraud Protection are all in the market, and they work. For social proof on Stripe's side, see Stripe Radar reviews on G2.

But none of them have matched Stripe's framing. Stripe is the only network-scale processor publicly describing its stack as a foundation model — one general-purpose model trained on the entire network, serving many tasks at once. Whether competitors eventually adopt the same architecture is almost beside the point; the positioning war is being won right now, and merchants evaluating providers in 2026 are hearing the foundation-model story first.

That framing also aligns with how CFOs and product leaders think about AI in 2026. They've all bought into the idea that a single large model, fine-tuned per task, beats a zoo of specialised models. Stripe is the first to ship that story convincingly in payments.

What This Means for Marketing Teams

You might be reading this and wondering what any of it has to do with marketing. Here's the connection.

Payments was the first operational surface where the answer to 'what should this system do next?' became a machine-learning question rather than a rules question. Fraud, authorisation, dispute prediction — all three started as rules engines, got replaced by ML, and have now consolidated under a single foundation model.

Marketing is roughly five years behind on the same curve, and the curve is accelerating. Content generation, channel selection, budget allocation, audience segmentation, creative iteration — all of these are currently handled by a zoo of specialised tools and humans wiring them together in spreadsheets. Exactly where payments was in 2018. The same consolidation is about to happen to marketing, and the winners will be the teams whose stack is already built around a single intelligent layer — like Anjin — rather than fourteen disconnected SaaS dashboards.

If Stripe's 2026 announcement proves anything, it's that foundation-model consolidation isn't a theoretical future state. It's already the competitive floor in one of the hardest, most adversarial operational categories in business. Marketing is next.

Anjin: The Marketing Operating System for the Foundation-Model Era

If Stripe Radar is what a foundation model does for payments, Anjin is the Marketing Operating System — what a foundation model does for marketing.

Anjin is a single platform that runs your marketing and distribution end-to-end. Content generation, SEO, campaign planning, channel distribution, brand consistency, performance tracking, reporting — all running inside one operating system, powered by agents that understand your brand the way Stripe's model understands payments.

What Anjin replaces:

• Your content agency (drafts, revises, publishes across channels)
• Your SEO consultant (optimises and tracks rankings continuously)
• Your paid media planner (briefs, tests, reports)
• Your distribution workflow (the 14 spreadsheets, Slack threads and Notion pages holding it all together)
• The £8–15k/month you're spending to coordinate it all

What Anjin does that none of them can:

• Runs 24/7. Your agency doesn't.
• Learns your brand voice in hours, not months.
• Ships campaigns the same day a news moment breaks.

The question every marketing leader should be asking in 2026 is the one Stripe already answered for payments: why are we still running a zoo of specialised tools when one intelligent layer can do the job better?

Sources: Stripe Radar, Stripe Payments AI, Stripe Engineering — Dynamic Radar Rules, Stripe — Payments Intelligence Suite, Radar for Fraud Teams, Stripe Docs — Advanced Fraud Detection, Stripe — 2025 State of AI and Fraud, G2 — Stripe Radar Reviews