A stolen card that buys API credits costs you twice. The compute is spent as soon as the tokens are served, and the dispute arrives weeks or months later with a fee and a mark against your card network ratios. Here is how the pattern works and where to stop it.
GuideHarry Johnston, NoResponse10 min read
Stolen cards buy AI credits because the credits arrive as soon as the payment clears, sell on through relays at a discount, and turn into compute that is spent long before the real cardholder disputes the charge. Card testing is the step before, when scripts run small payments or card saves through your top-up form in bulk to learn which stolen numbers still work.
This guide covers the case where an AI platform is the victim and the goods are tokens, which cost the platform money the moment a model serves them.
At a glance
AI credits are delivered as soon as a payment clears, resell easily and need no shipping address. By the time a bank reverses the payment, the platform has already paid for the compute.
Two Stripe Radar leads made the case in a September 2026 post on fraud at AI startups. AI startups, they wrote, "offer compute that's valuable and easy to resell, making them a particularly attractive target for fraud attacks." The same post says AI startups faced 4.3 times the attempted transaction fraud rate of startups overall in Q3 2025, falling to 2.6 times by Q1 2026. These are Stripe's own figures, from a company that sells the fix.
An account topped up with a stolen card can serve tokens through a relay, a site that sells unauthorised access to someone else's models, until the platform bans it. Our token resale guide covers how relays get their supply. TrustDecision, a fraud vendor, writes in its June 2026 article on token arbitrage that "With AI infrastructure, the financial loss is incurred the moment the tokens are processed."
The UK shows the same pattern from the cardholder's side. The Next Web reported in August 2026 that a Sussex businessman lost £14,244 when someone used his Metro Bank debit card to buy Claude credits, in charges "each between £90 and £200". Anthropic banned the user behind the fraud and refunded him directly.
Card testing is a run of small payments or card saves made to find out which stolen card numbers still work. On an AI platform it often hits the step where a new account adds a card to unlock credits or an API key.
Stripe's card testing guide says testers "use both card setup and payments" to check stolen numbers. It calls card setup the method testers prefer, because validations during setup don't usually appear on the cardholder's statement. Payments come second, in small amounts that cardholders are less likely to notice.
On a top-up flow, the signs Stripe lists look like this:
Stripe also warns that revenue from card testing "might look like good, new customers in your data". A wave of small top-ups can read as a good week until the disputes arrive.
Card testing also counts at the card network. Visa's Acquirer Monitoring Program fact sheet sets an enumeration threshold of 20% of authorisation attempts and 300,000 enumerated transactions a month. Stripe's monitoring programme notes say Visa places merchants over it in monitoring but does not fine them.
Radar and the network rules judge the payment, meaning the card, its history across Stripe and the issuer's answer. Neither sees what an account does with its credits after the payment clears.
Stripe's card testing guide says its recommended integrations, the Payment Element and Checkout, come with rate limiters, models and CAPTCHA triggers that Stripe switches on during an attack. They work better the more data your integration sends. The September post adds that Radar knows cards "that have been used fraudulently at other businesses, even when they're new to yours".
Custom velocity rules are where a platform tunes Radar for top-ups, and Stripe's rules reference says they depend on your Radar plan. The useful counters are in Stripe's list of supported rule attributes. Examples:
In our view, Radar cannot judge the second half of the loss. A stolen card with a clean history can pass every payment check, and the account then spends its credits on your most expensive model or hands its key to a relay. That pattern lives in your usage data, which Radar never sees.
A lost chargeback on consumed compute costs the top-up, the compute behind it and a dispute fee. A stolen-card dispute is also hard to win, because the real cardholder did not make the payment.
The timing works against you. Visa's Core Rules of 18 April 2026 give an issuer "120 calendar days from the Transaction Processing Date" to raise a card-absent fraud dispute. After a dispute opens, Stripe's guide to how disputes work says you have 7 to 21 days to respond, the issuer takes 60 to 75 days to decide, and the whole process can take 2 to 3 months. Stripe debits the disputed amount and a dispute fee when the dispute opens, and outside Mexico it does not return that fee even if you win.
Visa's compelling evidence rule needs the same card in "2 previous Transactions that the Issuer did not report as Fraud Activity", processed more than 120 calendar days earlier, with matching device or IP details. An account opened last week with someone else's card has no such history.
Early fraud warnings give you an earlier chance to act. Stripe says about 40% of Visa and Mastercard early fraud warnings become fraud disputes if you don't refund. A refund does not remove the fraud report, though, and only a reversal, usually within 2 hours of capture, prevents one.
Visa and Mastercard each place merchants with too many disputes or fraud reports under monitoring. Fines and remediation follow, and Stripe says a merchant that fails to comply can lose card acceptance.
Visa's programme divides fraud reports plus disputes by settled card-not-present transactions. The fact sheet's footnote says the Excessive Merchant threshold fell "to >=150bps in AP, Canada, EU, and U.S. regions on 1 April 2026", with at least 1,500 fraud reports and disputes a month. Stripe's monitoring notes add two details that matter for stolen cards. Visa counts early fraud warnings whatever their outcome, and a transaction that appears in both the fraud and dispute reports "will be counted twice".
We could not reach Mastercard's own rulebook, so the figures below come from Stripe's summary of Mastercard's programmes. A merchant enters Mastercard's Excessive Chargeback Merchant level at 100 to 299 chargebacks and a rate of 1.5% to 2.99%, measured against the previous month's sales. Fines start at $1,000 in months 2 and 3 and reach $100,000 from month 19. The separate Excessive Fraud Merchant programme applies only when, among other tests, 3-D Secure covers 10% or less of a merchant's Mastercard payments in non-regulated countries.
New platforms face one more programme. Stripe says Visa's Merchant Elevated Risk Program applies in Asia Pacific from October 2026 and identifies risk "early in a business' lifecycle". Visa does not publish its thresholds, and the consequences can include fraud disputes on payments past the normal deadline.
OpenCode's posts show the whole chain on one platform, from new accounts claiming a promotion to stolen cards and chargebacks. They also show payment controls stopping one part of it and usage data stopping another.
On 11 August 2026, dax, who builds OpenCode at Anomaly, explained what a fraudulent account was. People kept opening accounts for the initial promotion and would "issue credit card chargebacks to avoid actually paying anything". He added that "the cards are often stolen". On 25 August he posted that Stripe had blocked "$300M worth of fraud" for OpenCode in a few days. On 20 August he said a scheduled job now finds abuse in OpenCode's inference data.
Blocked payments never cost OpenCode compute. The cost came from accounts whose payments went through and were charged back after the tokens were used.
Make the first days of a new account cheap to lose. Tie spending to how long a card has been on your platform, keep expensive models behind a short wait, and link every card to every account that uses it.
Set a low first top-up and a daily spend cap for accounts and cards your platform has not seen before. Raise the limits as the card ages past your dispute window, or after a payment authenticated with 3-D Secure. In Radar, seconds_since_card_first_seen and amount rules can enforce this at payment time. Stripe's card testing guide also recommends limiting how many cards one account can add.
Serve your cheapest models first and open the expensive ones after a short wait or a cleared payment. A stolen card loses you least when it can only buy a small model's tokens. Stripe's monitoring notes add that an issuer does not report fraud on an authorisation you reverse before capture. A short gap between authorising a large top-up and capturing it gives your rules time to act.
One card on many accounts is the clearest signal in this guide. Check total_customers_for_card_weekly and email_count_for_card_weekly on every top-up, and treat a card that appears on a banned account as banned. Our free trial abuse guide covers the multi-account side, where the same operation, the group running the accounts, opens new ones after each ban.
3-D Secure usually moves liability for fraud disputes on authenticated payments to the card issuer. It also adds a step that some real customers will abandon. Stripe's 3-D Secure documentation says liability shift "isn't guaranteed for any specific payment or dispute". It does not apply when a card doesn't support 3-D Secure, and early fraud warnings still arrive. Stripe's monitoring notes add that Mastercard counts a chargeback towards its thresholds regardless of liability shift, while a high 3-D Secure share keeps you out of its fraud programme. We would request 3-D Secure on the first top-up from a new card, on any top-up above your new-account cap and on cards Radar rates as elevated risk. Stripe can ask the issuer to skip the challenge screen where possible, which keeps the cost to real customers down.
You need your Stripe dashboard and your usage logs for these.
Payment checks assess the card. NoResponse assesses the account and its usage.
Our token fraud guide shows where stolen cards sit among the other ways tokens leak from a platform. If disputes on top-ups are rising, Book a demo.
Every token an AI platform serves runs on compute someone pays for. Token fraud is when the person using it is not the person you sold it to. This guide maps the four ways that happens and links to a guide for each.
GuideEvery free credit on an AI platform is compute someone pays for. Score each sign-up before it gets credits, stage the credits and keep the card for accounts that raise a signal.
GuideA resold account looks like one customer and behaves like a hundred. This guide sets out the signals in your own logs that give it away, how to confirm resale without hurting real customers, and what to do once you find it.
GuideYou will rarely win one when the card was stolen. The real cardholder did not authorise the payment, and Visa's compelling evidence rule needs at least two earlier undisputed payments on the same card, which a new account does not have.
Stripe says its recommended integrations block much card testing, and results depend on the data your integration sends. Custom velocity rules depend on your Radar plan.
