Zenpanik
The coat check: card tokens, vaults, and who can ask for the number backBy Pavel Nikolov, November 16, 2026

Part 5: Guarding the counter

Part 5 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.

The counter is a set of rules

A vault is as safe as its rules for redeeming a token. I would write those rules as three questions asked of every request: who is asking, where the number will be sent, and how many numbers this caller has already had today. Figure 12 runs three requests past them.

Request to the vaultCounter rulesOutcomeWho is askingWhere it is sentHow many todayCharge, to a processor on the listCaller: checkout serviceNumber releasedSame token, to an unknown hostCaller: checkout serviceRefusedShow 5,000 full numbersCaller: support tool, limit 50 a dayRefused and flagged
Figure 12. Three requests at the counter. A charge sent to a processor on the approved list gets the number put back. The same token sent to an unknown host is refused, and so is a bulk request far above the caller's daily limit. Animated. My proposal, with invented callers and limits.

Seven practices

1. Remove the number before you protect it

Start by finding every place a full card number still exists: databases, application logs, analytics data, support tickets, call recordings, exports and backups. Delete what you do not need. This is the via negativa step from Part 4, and the PCI council gives the same advice for hashed numbers, which is to remove a number that will never be needed again. If you keep a fingerprint for matching, make it a keyed hash. Keep it away from any truncated copy of the same number, because the council requires extra controls when both sit in one environment.

2. Capture in a hosted field and watch the page around it

Let the vault or the processor serve the card input, so the number goes from the shopper's browser to the vault and never reaches your servers. Then treat the page that holds the frame as part of the payment system. Keep an inventory of every script on it, including analytics. Check that the real payment frame is present and has the size you expect, as Report URI recommends. Convertr was skimmed through an analytics file, and the fake Stripe forms worked by hiding the real form.

3. Redeem tokens only in transit, to named destinations

Use the proxy pattern from Part 2, so that a token is turned back into a number only on its way to a processor and never inside your network. I would give each outbound route one destination, and each calling service its own credential, a daily limit and an alert when it gets close. If a staff tool can show full card numbers, I would remove that feature or put a named approval in front of it. A vault credential deserves the same care as the signing authority I wrote about in Two clocks.

4. Use network tokens for saved cards, behind a token you own

For cards on file, store network tokens and send a cryptogram with every payment the customer starts. A stolen file of merchant-restricted tokens is of little use elsewhere. Keep your own vault token in front of the network token, the dual-tier arrangement from Part 1, so that you can change processors and your systems hold one reference per customer whatever sits behind it. Budget for the cost. Since April 2026 token services are part of a fee on online authorizations.

5. Guard the ticket desk

If you issue cards or run a wallet, the enrollment check is your most important control. Approve new tokens inside your own app with the customer logged in, and stop treating a texted code as proof, as Krebs's sources recommend. Use the age of a token as a risk signal, as Visa advises, because the gangs wait about a week before spending.

6. Give agents scoped tokens, never numbers

An AI agent that buys things should hold a token limited to one seller, one amount and a short time. Stripe's Shared Payment Tokens are built this way, and Mastercard says its Agentic Tokens keep an agent inside the limits the user sets. ChatGPT Wallet asks for a face or fingerprint check on every purchase at launch. OpenAI expects to let users pre-approve purchases within spending limits later, The Paypers reports. When that arrives, set the limits low and keep a person on anything large or hard to reverse. I would manage a token handed to an agent like any other credential, with an owner, a limit and a log.

7. Limit what one vault and one caller can give up

This is the first rule in Taleb's letter, to split sources of harm as much as we can, provided they do not interact. Per-caller limits cap what a stolen credential can take. Separate vaults for separate regions or business lines cap what one breach can take. Get the right to export your card data written into the vault contract and test the export, so that leaving is possible. The FTC's allegations against Mastercard describe what can happen when only one party can turn a token back.

Where it is takenVault tokensHosted field and page controlsNetwork tokens with cryptogramsIn-app approval at enrollmentRedeem rules and limitsSplitting and exit rights
The lobby: checkout page
The rack: stored numbers
The ticket desk: enrollment
The counter: redeeming tokens
Concentration in one vault
StrongPartialNot covered
Figure 13. Which control covers which of the places from Part 4, as I assess it. No column covers every row, and the two token columns leave the ticket desk open.

Network tokens cover the lobby only in part. A shopper who pays from a phone wallet types nothing, while one who types a card gives a skimmer the real number whatever the merchant does with it afterwards. They also help at the counter when the vault stores network tokens, because what comes back from a redeemed token is still restricted to one merchant.

Where to start

Number safety

List every system that can see a full card number, including logs, support tools and analytics. Delete what you can. Put a hosted field in the checkout and an inventory on the scripts around it. Move saved cards to network tokens with cryptograms.

Counter safety

List every caller that can redeem a token or approve a new one. Give each a named destination, a daily limit and an alert. Approve wallet enrollments in your own app. Give AI agents tokens with a seller, an amount and an expiration time, and keep a person on large purchases.

Sources

  1. Tokenization (data security), Wikipedia
  2. Visa Q4 2026 earnings call transcript, October 27, 2026, Stock Analysis
  3. Revolve Group, Form 8-K, October 29, 2026, SEC
  4. Revolve puts data breach victim count at 1.5 million, SecurityWeek
  5. Fandango Media, Notice of Data Breach, California Attorney General
  6. Card security code, Wikipedia
  7. Hardware wallets, randomness, and the vaults behind every device, Zenpanik
  8. Two clocks: quantum breaks the math, AI breaks everything around it, Zenpanik
  9. Card scheme, Wikipedia
  10. PCI DSS Tokenization Guidelines, Information Supplement, August 2011, PCI Security Standards Council
  11. FAQ 1326: How does PCI DSS apply to EMVCo Payment Tokens, PCI Security Standards Council
  12. VGS surpasses 10 billion tokens stored, Very Good Security
  13. Visa issues 10 billionth token, Visa, June 4, 2024
  14. Mastercard advances Europe's checkout transformation on the road to 2030, Mastercard
  15. Fandango says hackers stole payment card details from checkout pages, BleepingComputer
  16. OpenAI launches ChatGPT Wallet for agentic payments in the US, The Paypers
  17. Visa Merchant Business News Digest, Visa
  18. Mastercard touts the importance of agentic tokens, American Banker
  19. Brazilian Magecart group hijacks Convertr checkouts via outdated Plausible install, Sansec
  20. Malicious AI agents steal 600K credit cards, infect 100+ sites with skimmers, BleepingComputer
  21. Fake payment fields: skimmers that replace a hosted element, Report URI, August 24, 2026
  22. Spring 2026 Biannual Threats Report, Visa
  23. April 2026 US network fee updates, CMSPI
  24. Legacy Stripe API exploited to validate stolen payment cards in web skimmer campaign, The Hacker News
  25. Guidance for PCI DSS e-commerce requirements effective after 31 March 2025, PCI Security Standards Council
  26. How phished data turns into Apple and Google wallets, KrebsOnSecurity
  27. SP 800-38G Rev. 1, second public draft, Methods for Format-Preserving Encryption, NIST
  28. FTC orders an end to illegal Mastercard business tactics, Federal Trade Commission, December 23, 2022
  29. Capital One announces data security incident, Capital One, July 29, 2019
  30. What are network tokens and how do they work, Evervault
  31. Target confirms unauthorized access to payment card data in US stores, Target, December 19, 2013
  32. Apple Pay security and privacy overview, Apple
  33. Sam Boboev, Deep Dive: How to add data intelligence to a multi-PSP payment stack, Fintech Wrap Up, July 19, 2026
  34. Sam Boboev, Deep Dive: The hidden liability of agentic commerce, Fintech Wrap Up, March 29, 2026
  35. Payment card number, Wikipedia
  36. Luhn algorithm, Wikipedia
  37. FAQ 1492: PAN masking and truncation with 8-digit BINs, PCI Security Standards Council
  38. Cryptographic hash function, Wikipedia
  39. NVIDIA RTX 4090 benchmark, hashcat 7.0.0, hashcat forum
  40. FAQ 1091: Acceptable formats for truncation of primary account numbers, PCI Security Standards Council
  41. HMAC, Wikipedia
  42. FAQ 1308: How can an entity ensure that hashed and truncated versions cannot be correlated, PCI Security Standards Council
  43. FAQ 1089: How can hashing be used to protect PAN, PCI Security Standards Council
  44. Encryption, Wikipedia
  45. Inbound route, VGS workshop for AWS
  46. Proxy server, Wikipedia
  47. HTTPS proxy, VGS documentation
  48. Aliases, VGS documentation
  49. Comparing Evervault to Basis Theory, Basis Theory
  50. Choosing the right tokenization solution for your business, IXOPAY
  51. Format-preserving encryption, Wikipedia
  52. NIST requests comments on Draft SP 800-38G Revision 1, NIST, February 2019
  53. The what, why and how of EMV Payment Tokenisation, EMVCo
  54. One year in, Mastercard's checkout transformation gains ground across Europe, Mastercard
  55. Agentic commerce momentum, Mastercard
  56. Shared payment tokens, Stripe documentation
  57. Two clocks, Part 5: Posture at machine speed, Zenpanik
  58. Biannual Threats Report, Spring 2025, Visa Payment Ecosystem Risk and Control
  59. Two clocks, Part 3: The AI clock, Zenpanik
  60. Nassim Nicholas Taleb, letter to Stewart Brand, Artangel Longplayer Letters, July 3, 2013
  61. Nassim Nicholas Taleb, On the difference between a large stone and a thousand pebbles, Antifragile draft chapter
  62. The Black Swan: The Impact of the Highly Improbable, Wikipedia
Back to Part 1