Where the term comes from
Is every token stored somewhere? Not always. The PCI Security Standards Council, which writes the card industry’s data rules, splits reversible tokens into two families in its April 2015 product guidelines.
For reversible cryptographic tokens, “the PAN is never stored; only the cryptographic key is stored.” In vendor language, that is the vaultless design. For reversible non-cryptographic tokens, the card number comes back “only by a data look-up in a card data vault”, for example from a pre-generated table of random values.
A hybrid also exists, the guidelines note, in which a cryptographic token is still stored in a vault, and then both sets of rules apply. Security vendor Thales sells the vault and the key-only designs, listing “Vaulted and Vaultless solutions” on its tokenization page.
Vault or key
What does a tokenization vault actually hold? Visa’s fact sheet shows its token service looking up the card number “stored in the token vault” on every payment, the model described under network tokenization.
| Token vault | Vaultless | |
|---|---|---|
| What is stored | Each original value beside its token | The key only |
| Link between token and value | None, apart from the table | Mathematical, through the algorithm |
| What an attacker wants | The vault, “the most attractive target” in PCI’s 2011 guidance | The key, whose theft could expose “all current and future tokens” made with it, per the same guidance |
| Size over time | At least one record for each person in the data, per the EDPB | Fixed, since the key does not grow with the data |
The size row comes from the European Data Protection Board’s 2025 guidelines, which add that lookup tables are themselves personal data. Controllers, the board says, must weigh guarding that large table against the protection it gives from cryptanalytic attacks.
The algorithm under it
Key-based tokens can be built with format-preserving encryption, which keeps a 16-digit number at 16 digits. Thales says format preservation “avoids the need to change your database schema.”
The US standards body NIST approved two such methods, FF1 and FF3, in SP 800-38G of March 2016. Researchers then found weaknesses when the number of possible inputs is small.
NIST’s second public draft of a revision, published February 3, 2025, says “the encryption method FF3 is no longer specified” and bans floating-point arithmetic in FF1 code. Comments closed on April 4, 2025. When this entry was checked on November 12, 2026, the revision was still listed as a draft.
The mistake
The trap is assuming vaultless means as strong as any encryption. Google warns in its cloud documentation that format-preserving encryption “provides fewer security guarantees” than deterministic encryption with AES-SIV, and that short values drawn from a small alphabet can be reverse engineered.
Ask a vendor which algorithm it uses, whether FF3 is still in the product, and how the key is stored and rotated, since in a vaultless design the key is the whole secret.