Part 4: Where the number went
Part 4 of 5 in a deep dive on card tokens, token vaults and where the card number went. Start with Part 1 if you are new to the series.
Four places a card can still be taken
Put the two kinds of token together and the back room is closed. A merchant that uses both stores nothing a stranger can spend. The cases from the past two years show where the thefts went. I count four places, and I have public cases for three of them.
1. The lobby: before the number becomes a token
Every saved-card token starts with a person typing a number into a web page. Web skimming is the theft of card data by malicious code that runs inside the checkout page. The code reads what the customer types and sends a copy to the attacker.
Fandango
Fandango's notice says an unauthorized party could access the card number, security code, expiration date, billing ZIP code and email address from certain purchases on its checkout page between October 8 and 11. BleepingComputer describes the cause as skimmer code injected into the page. That is the outlet's description, and the notice does not name the method. Fandango sells tickets for about 31,000 US movie screens and has not said how many customers were affected.
Convertr
Analytics software counts a site's visitors and records what they do. The security company Sansec found a skimmer in an analytics file on the servers of Convertr, a Brazilian platform that runs the online shops of dozens of fashion brands. Every Convertr checkout loaded that file. The skimmer read the card number, expiration date, security code, name and taxpayer number from the payment form and from the requests the checkout sent to its payment service. Sansec detected it on October 24 and Convertr removed it on October 27. Sansec suspects the way in was an analytics tool that had not been updated since a fix came out in May, and says it has not seen the server logs.
Fake Stripe forms
The security companies Jscrambler and Source Defense tracked a campaign, reported by The Hacker News, that had run since at least August 2024 and reached an estimated 49 merchants. The script hid the real Stripe payment form, drew a copy in its place, and checked each stolen card against an old Stripe interface so that only valid cards were sent to the attackers.
The third case shows the limit of a hosted field. A hosted field is a card input that the vault or the processor serves directly. It sits in a frame. A frame is a small window inside the page whose contents come from another site, and the browser stops the merchant's page from reading what is typed into it. Report URI, a web security company, described the gap in August 2026. The browser's barrier only blocks reading. A script on the surrounding page can still remove the frame, make it invisible or tiny, or paint its own card fields over it. The attackers in the Stripe case hid the real form and collected the numbers in a copy of their own.
PCI DSS has had rules for this since March 31, 2025, when requirements 6.4.3 and 11.6.1 became mandatory. They require that the scripts on a payment page are authorized, checked for integrity and monitored for tampering. A merchant has to know every script on its payment page and notice when one changes.
2. The shop's own rack: numbers that were never checked in
Tokenization protects only the numbers that are tokenized. Revolve's saved cards and the 600,000 records in Gambit's report, both covered in Part 2, were card numbers that an intruder on the shop's platform could collect. Gambit's report adds a price. At about $25 a target, checking large numbers of shops for stored card numbers is cheap.
3. The ticket desk: getting a token issued
Provisioning is the industry's word for issuing a token to a device or a merchant. A network token is as trustworthy as the check made at that moment. Phishing is tricking a person into typing secrets into a fake page or message. Visa defines provisioning fraud as fraud within seven days of a token being activated, mostly on tokens bound to a device, and says the usual method is phishing for the one-time passcode the bank sends.
The security journalist Brian Krebs documented how it is done with phishing kits sold by vendors in mainland China. A message about a package or a road toll sends the victim to a fake page that collects the card details. The criminals add the card to Apple Pay or Google Wallet on a phone they control. The bank sends a one-time code to its customer, and the fake page asks the victim to type that code in. The token that results is real. Krebs reports such phones being sold in batches of ten or more, each loaded with several wallets.
The two sources agree on timing. Krebs's sources say the gangs now wait about seven to ten days before using a phone, down from 60 to 90. Visa's report for the second half of 2024 found fraud in the first seven days after activation down 29% as criminals waited longer, and advised banks to treat the age of a token as a risk signal.
The spending then looks legitimate. Krebs describes an Android app, first documented by ThreatFabric in November 2024, that relays a phone's tap over the internet to an accomplice standing at a shop terminal somewhere else. The payment that reaches the terminal is a valid one.
The fixes Krebs's sources propose are all at the ticket desk. Make the customer approve a new wallet inside the bank's own app, as some European and Asian banks already do. Stop using a texted code as the default proof. Tie the approval to the device that is asking for the token.
4. The counter: asking for the number back
I have no public card case for the fourth place. A vault redeems tokens for callers it trusts. VGS's documentation names two routes. An alias can be swapped for the real value in transit to a third party, or fetched with a request to the vault's API. A credential is the secret a program presents to prove who it is. As I read the two routes, the number goes to whoever holds the right credential or controls the server behind the right route. Someone who takes over one of those callers could then collect numbers one request at a time without touching the vault's storage. That is my inference, and I have no card case that shows it.
The closest documented cases come from outside card vaults. At Capital One the intruder's access also allowed data to be decrypted, and the tokenized fields stayed protected. The bank says it tokenizes them with a method and keys separate from its encryption. At Binance, in a case I covered in Two clocks, attackers reached the service that signs withdrawals and it signed about $6.34 billion for them. I think the interface that turns tokens back into numbers is the same kind of service. The PCI council called the data vault the most attractive target for attackers in 2011.
One vault in place of a thousand databases
Tokenization also changes how the risk is spread. Before it, card numbers sat with every merchant that kept them. After it, a growing share sits with vault operators, which are the card networks' token services, the processors and independent companies such as VGS. Over half of Visa's available credentials now have a token, and VGS reports more than 10 billion tokens in its vault.
A small model shows what that does. Suppose each of N independent holders is breached with probability p in a year and holds n cards. Call the number of breaches in a year B. B follows the binomial distribution, the one that counts successes in N independent yes-or-no trials. Its mean is Np and its standard deviation is √(Np(1−p)). The standard deviation measures how far a typical year falls from the mean. The yearly loss is n × B, so its mean is Npn and its standard deviation is n√(Np(1−p)).
Take 1,000 merchants with 100,000 cards each and a 1% chance of a breach. The expected loss is 1,000 × 0.01 × 100,000 = 1 million cards a year. The standard deviation is 100,000 × √(1,000 × 0.01 × 0.99) = 100,000 × √9.9, about 315,000 cards. Most years look alike.
Now move the same 100 million cards into one vault with the same 1% chance. The expected loss is still 1 million cards a year. The standard deviation is 100 million × √(0.01 × 0.99), about 9.9 million cards. That is roughly 32 times larger, and 32 is √1,000. In 99 years out of 100 nothing is lost, and in the other year everything is. These numbers are illustrative. I chose them to show the shape.
I assume a real vault is much better defended than a typical merchant, so its yearly chance of a breach should be far below 1%. I have no measured rate for either. At 0.1% the expected loss falls to 100,000 cards a year, a tenth of the merchants' figure. The standard deviation is still about 3.2 million cards, ten times theirs. Concentration lowers the average and widens the spread, and a comparison of averages shows only the first effect.
For scale, Target's 2013 breach affected about 40 million card accounts. A vault with 10 billion tokens holds 250 times as many records, though one card can have many tokens and a token can stand for data other than a card.
A Taleb reading
Nassim Taleb's test for fragility applies to this directly. In his 2013 letter to Stewart Brand he writes that anything fragile is harmed more by one stressor of intensity X than by k stressors of intensity X/k, and gives the image of a large stone against the same weight in pebbles. In mathematical terms the harm is a convex function of the size of the shock. A function is convex when its graph curves upward, so each added unit of input adds more output than the one before. Here that means doubling the shock more than doubles the damage.
I think the harm from a card breach is convex in its size, because the response has fixed capacity. My assumption is that a bank can absorb a few hundred thousand reissued cards as routine work, and that a hundred million at once would overload card production, call centers and fraud teams at the same time.
Put an illustrative number on it. If harm grows as size1.5, one event of 100 million cards does (108)1.5 = 1012 units of harm. A thousand events of 100,000 cards do 1,000 × (105)1.5, about 3.2 × 1010. The cards are the same and the harm is about 32 times larger. The exponent is mine and shows direction only.
Taleb's third rule in the letter is to decentralize variations. He writes that nature produces many small, independent variations while human systems produce fewer and larger ones, and that a decentralized system is more effective than a command-and-control one because its errors are decentralized and do not spread. Card numbers scattered over a thousand merchant databases lose on the average if a vault is as well defended as I assumed above. On this one measure they do better, because no single failure exposes everything.
The record so far favors the vaults. I found no public report of a breach of a card network's token service or of VGS's vault, and Visa's numbers for token fraud are falling. Taleb's turkey is the caution against reading too much into a clean record. In The Black Swan the turkey is fed every day, and every feeding raises its confidence that the farmer cares for it, until the day it is killed for Thanksgiving. Twelve years without a public report of a vault breach is good evidence about ordinary years. It says little about a design whose worst case is every card at once.
Taleb also connects fragility to time. In the letter he names the mechanism the Lindy effect and says that time is good at showing what breaks and what does not. My reading for cards is this. The weaknesses of the 16-digit card number are well mapped, as Part 2 showed. Network tokens date from 2014, so time has had twelve years to test them, and they already carry most of Visa's online volume.
The best case for tokenization is also in the letter, under the name via negativa. Via negativa is his name for acting by removal. His point is that the effects of adding something new to a complex system are hard to predict, while the effect of removing something that was not historically part of it can be predicted. A merchant that stops storing card numbers has removed something, and the result is predictable, because there is nothing left on its servers to steal. A merchant that adds a vault, two proxies, a hosted field and a token service has added five things. I read the safety as coming from the removal.
Where this argument stops
Three things keep this argument in proportion. First, the averages favor tokens. Fraud on tokens held in devices is falling in Visa's data, and the breaches of 2026 in this piece took numbers that were never tokenized or had not been tokenized yet. Second, I am arguing about the tail of the distribution. The tail is the set of rare, extreme outcomes, and no observed event sits in it. I cannot test the argument against data, because the only data that would confirm it is a vault breach.
Third, the alternative to a few large vaults is thousands of small stores of card numbers. I assume many of them would be run by companies with little security staff. That is the situation tokenization was introduced to end.
What I take from Taleb is to use vaults and to limit how much any one vault and any one caller can give up. Part 5 lists the controls I would put in place.
Continue to Part 5: Guarding the counter