FAST format-preserving cipher for Python

A pure Python implementation of the FAST cipher, a format-preserving encryption (FPE) scheme designed for tokenizing API keys, credentials, and other structured secrets.

For prefix-based tokens (GitHub, AWS, Stripe, etc.), encrypted output keeps the exact same format (length, prefix, character set) as the originals, so they pass through systems that validate token formats. Heuristic tokens (Fastly, AWS secret keys) are wrapped in a tagged marker since they have no distinguishing prefix.

Fully interoperable with the C, Zig, Go, and JavaScript implementations.

Installation

pip install fast-cipher

Or with uv:

uv add fast-cipher

Quick start

The most common use case is encrypting tokens and API keys found inside a block of text. TokenEncryptor handles scanning, encrypting, and decrypting automatically:

import os
from fast_cipher.tokens import TokenEncryptor

key = os.urandom(32)  # AES-128, AES-192, or AES-256
encryptor = TokenEncryptor(key)

text = """
Here are the credentials for the staging environment:
GitHub token: ghp_ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghij
AWS access key: AKIAIOSFODNN7EXAMPLE
Stripe key: sk_live_ABCDEFGHIJKLMNOPQRSTUVWXab
"""

encrypted, mappings = encryptor.encrypt(text)

encrypt() returns a tuple: the encrypted text and a list of TokenMapping objects that record which tokens were found and what they became. Each mapping has plaintext, ciphertext, and pattern_name fields, which is useful for provenance tracking (only reversing tokens you actually encrypted).

For prefix-based tokens the result still looks like valid tokens: same prefixes, same lengths, same character sets, but the secret parts have been replaced with ciphertext. Decryption restores the original text exactly:

decrypted = encryptor.decrypt(encrypted)
assert decrypted == text

Tweaks

A tweak is optional context data that gets mixed into the encryption. The same plaintext encrypted with different tweaks produces different ciphertext, which is useful for binding tokens to a specific user, session, or tenant:

enc_alice, _ = encryptor.encrypt(text, tweak=b"user-alice")
enc_bob, _ = encryptor.encrypt(text, tweak=b"user-bob")

assert enc_alice != enc_bob

# Each can only be decrypted with the matching tweak
assert encryptor.decrypt(enc_alice, tweak=b"user-alice") == text
assert encryptor.decrypt(enc_bob, tweak=b"user-bob") == text

Filtering by token type

If you only want to encrypt certain kinds of tokens and leave the rest as-is, pass a types list with the pattern names you care about:

text = "ghp_ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghij and AKIAIOSFODNN7EXAMPLE"

encrypted, _ = encryptor.encrypt(text, types=["github-pat"])
# The GitHub token is encrypted, but the AWS key is untouched

Supported token types

The following patterns are detected and encrypted out of the box:

Heuristic patterns don't rely on a prefix. They look for strings with high entropy and mixed character classes, which is how secrets like Fastly tokens and AWS secret keys are typically formatted. Because there is no distinguishing prefix, encrypted output is wrapped in an [ENCRYPTED:<name>] marker. decrypt() will attempt to unwrap anything matching that marker pattern, so avoid feeding text containing literal [ENCRYPTED:...] strings that were not produced by encrypt().

Custom token patterns

You can register your own patterns for tokens that aren't covered by the built-in set. A SimpleTokenPattern works for anything that has a fixed prefix followed by a body with a known alphabet:

from fast_cipher.tokens import ALPHANUMERIC, TokenEncryptor
from fast_cipher.tokens.types import SimpleTokenPattern

my_pattern = SimpleTokenPattern(
    name="myapp-api-key",
    prefix="myapp_",
    body_regex="[A-Za-z0-9]{32}",
    body_alphabet=ALPHANUMERIC,
    min_body_length=32,
)

key = os.urandom(32)
encryptor = TokenEncryptor(key)
encryptor.register(my_pattern)

text = "key: myapp_ABCDEFGHIJKLMNOPQRSTUVWXYZabcdef"
encrypted, _ = encryptor.encrypt(text)
decrypted = encryptor.decrypt(encrypted)
assert decrypted == text

Registered patterns take priority over built-in ones.

The available alphabets are DIGITS, HEX_LOWER, ALPHANUMERIC_UPPER, ALPHANUMERIC_LOWER, ALPHANUMERIC, BASE64, and BASE64URL. You can also create your own with Alphabet(name="my-abc", chars="abc...").

Low-level cipher

TokenEncryptor is built on top of FastCipher, which you can use directly when you need format-preserving encryption for arbitrary data. It works on sequences of integers in a given radix (base).

For example, to encrypt an 8-digit decimal number:

from fast_cipher import FastCipher, calculate_recommended_params

params = calculate_recommended_params(radix=10, word_length=8)
key = os.urandom(32)
cipher = FastCipher(params, key)

plaintext = [1, 2, 3, 4, 5, 6, 7, 8]
ciphertext = cipher.encrypt(plaintext)

# Result is still 8 digits, each between 0 and 9
assert len(ciphertext) == 8
assert all(0 <= d < 10 for d in ciphertext)

decrypted = cipher.decrypt(ciphertext)
assert decrypted == plaintext

For raw bytes, use radix 256 with the encrypt_bytes/decrypt_bytes convenience methods:

params = calculate_recommended_params(radix=256, word_length=16)
cipher = FastCipher(params, key)

ciphertext = cipher.encrypt_bytes(b"sensitive data!!")
plaintext = cipher.decrypt_bytes(ciphertext)
assert plaintext == b"sensitive data!!"

Cleanup

When you're done with an encryptor or cipher, call destroy() to invalidate the instance:

encryptor.destroy()

After destroy(), any further calls to encrypt() or decrypt() will raise a RuntimeError. The method overwrites key references with zero-filled placeholders and clears internal state, but Python's garbage collector may retain copies of the original key bytes in memory. For applications that require guaranteed memory scrubbing, use a C-level implementation instead.

Cross-implementation compatibility

Ciphertext produced by any FAST implementation (C, Zig, Go, JavaScript, Python) can be decrypted by any other, as long as the key, radix, word length, and tweak match. This library is tested against the Go test vectors and cross-validated against the JavaScript implementation.