NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI, #330 most downloaded on PyPI
Self-contained cryptographic library for Python, classical and post-quantum
Latest stable
4.6.0
Last release today
22 Nov 2026
Ships unpredictably
gaps range from 9 days to 1.4 years
Nearly every release is documented
notes for 60 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
12 years old
68 releases · first in 2014
Drag across the ruler to compare, or tap a release to open it.
New module Crypto.PublicKey.SLHDSA , for SLH-DSA keys (FIPS 205), with all twelve parameter sets (SHA2 and SHAKE, categories 1, 3 and 5, small or fast
Crypto.PublicKey.SLHDSA, for SLH-DSA keys (FIPS 205), with allSubjectPublicKeyInfo (also from certificates),OneAsymmetricKey (also encrypted), in DER or PEM.Crypto.Signature.slhdsa, for SLH-DSA and HashSLH-DSA signatures (FIPS 205),construct() (which checks the private key)threads, to compute the XMSS and FORS trees on several threads.s parameter sets is 3.5x to 5x faster.Crypto.PublicKey.SLHDSA also supports the six parameter sets of the initial publicSLH_DSA_SHA2_128_24: smaller signaturesthreads). The draft may still change: their support is--include-very-slow-tests, for the tests marked very_slowCrypto.Cipher.AEGIS, for the authenticated ciphers AEGIS-128L andCrypto.Signature.mldsa.new(), the context can only be passed as a keywordcontext=), as for Crypto.Signature.slhdsa.new().ECC.construct(), and so ECC.import_key(), accepted the point at infinity04 followed by zero bytesSubjectPublicKeyInfo). ECDSA signatures for such a key can be forged withoutValueError. Thanks to Zulfareen Shaikh.DSA.construct(), and so DSA.import_key(), accepted public keysq (for instance, p - 1).y = 1, a signature can be forged on any message without any secret.ValueError (NIST SP 800-89, 5.3.1). Thanks to Zulfareen Shaikh.decrypt_and_verify() method of GCM, CCM, EAX and SIV, the output bufferINTEGER took time quadratic in its length, so a craftedRSA.import_key) kept the CPU busy for minutes.OBJECT IDENTIFIER had the same problem: itsValueError. Thanks to Mrmaxmeier.Integer objects.Your coding agent can read these notes before it upgrades. Set up the MCP server →
Added the mode MODE_GCM_SIV to Crypto.Cipher.AES (AES-GCM-SIV, RFC 8452), for authenticated encryption that is resistant to the reuse of a nonce.
MODE_GCM_SIV to Crypto.Cipher.AES (AES-GCM-SIV, RFC 8452),MODE_XTS to Crypto.Cipher.AES (IEEE 1619, NIST SP 800-38E),bytes messages.export_key of RSA and DSA keys raises ValueError for a private key in the 'OpenSSH' format,passphrase.key.public_key().export_key('OpenSSH').binascii.a2b_base64, which looks up each character in a table, and could leakAdded the elliptic curve secp256k1 to Crypto.PublicKey.ECC : key generation, import and export, ECDSA with Crypto.Signature.DSS , and ECDH with Crypto
Crypto.PublicKey.ECC: key generation,Crypto.Signature.DSS, and ECDH withCrypto.Protocol.DH.'sec1' to Crypto.Signature.DSS.new, for randomized ECDSA on all Weierstrass'fips-186-3' does not accept secp256k1, which is not a curve of that standard.low_s to Crypto.Signature.DSS.new, for ECDSA signaturess, and verifying rejects a high s.ECC.import_key raised TypeError instead of ValueError for an OpenSSH public keyECC.import_key accepted an OpenSSH key (public or private) whose curve did not match its key type.unseal() consumed a sequence number, so all laterunseal()seal() that raised an exception also consumed a sequence number,CTR mode is 2.5-3.5x faster on large data (and GCM about 2x), on all platforms.
Integer.random_range() is about 2x faster when the bounds are public,Integer objects.Crypto.Cipher.PKCS1_v1_5) performs implicit rejection,draft-irtf-cfrg-rsa-guidance,decrypt() is called without a sentinel (which is now optional):sentinel, decrypt() behaves as before.chacha20_counter_test.json).sentinel returned for ciphertexts with incorrect padding.decrypt() without sentinel (implicit rejection) is nowdecrypt() never returned a sentinel that is not a byte stringNone), and it returned an empty byte string instead.Crypto.Util.asn1.DerObjectId accepted an ASN.1 OBJECT IDENTIFIER where an arcNew module Crypto.PublicKey.MLKEM , for ML-KEM keys (FIPS 203). It is based on the portable C code of mlkem-native <https://github.com/pq-code-package
Crypto.PublicKey.MLKEM, for ML-KEM keys (FIPS 203).mlkem-native <https://github.com/pq-code-package/mlkem-native>_ (v2.0.0),Crypto.PublicKey.Hybrid, to model hybrid keys (post-quantum + traditional).Crypto.Protocol.HPKE accepts ML-KEM keys and X-Wing keys,New module Crypto.PublicKey.MLDSA , for ML-DSA keys (FIPS 204). It is based on the portable C code of mldsa-native <https://github.com/pq-code-package
Crypto.PublicKey.MLDSA, for ML-DSA keys (FIPS 204).mldsa-native <https://github.com/pq-code-package/mldsa-native>_ (v2.0.0), with the Keccak code of PyCryptodome. On CPUs with AVX2 and BMI2, a version compiled for them and with 4-way SHA-3 is used: compared to the portable version, key generation and verification are 30-45% faster, signing 15-30%.Crypto.Signature.mldsa, for ML-DSA and HashML-DSA signatures (FIPS 204), hedged (default) or deterministic, with an optional context. The new MLDSAMuHash computes the message representative mu (external mu), to sign or verify a large message in pieces.digest() in the modules of the SHA-2 and SHA-3 hashes, to hash a whole message with a single call, like Crypto.Hash.KangarooTwelve.digest(). For messages up to 1 KiB, it is 1.5x to 3.4x faster than new() followed by digest().Crypto.IO.PKCS8.wrap() has a new parameter public_key, to create a OneAsymmetricKey version 2 (RFC 5958) with the field publicKey. Crypto.IO.PKCS8.unwrap() returns an UnwrappedKey: still a tuple (oid, private_key, params),public_key (the field publicKey, if present).Crypto.PublicKey.ECC.import_key() checks that the field publicKey of a PKCS#8 OneAsymmetricKey version 2 (RFC 5958), if present, matches the private key. Before, it was ignored.The Windows wheels for x86-64 are built with clang-cl (compatible with MSVC), so the code for CPUs with AVX2, BMI2 and ADX is also used on Windows: SH
clang-cl (compatible with MSVC), so the
code for CPUs with AVX2, BMI2 and ADX is also used on Windows: SHA-3 and the other
Keccak-based functions, KangarooTwelve, ChaCha20, Salsa20, and the big integers of RSA,
DSA and the NIST curves. To build from source with clang-cl, set the environment
variable PYCRYPTODOME_CLANG_CL (see the installation instructions).Major release, with some breaking changes. See Changelog.
Major release, with some breaking changes. See Changelog.
Remove support for Python 2.7 and 3.8. The minimum supported version is now Python 3.9.
The pycryptodomex package (Cryptodome namespace) is no longer released (except for important security patches, but still on the 3.x branch). The library is only available as pycryptodome, under the Crypto namespace.
The self-tests are now run by pytest, which must be installed for python -m Crypto.SelfTest to work. The test command of setup.py is removed: install the package (e.g. with pip install -e .) and run python -m Crypto.SelfTest instead.
In Crypto.Protocol.KDF.SP800_108_Counter(), the label can no longer contain zero bytes, while the context now can (for instance, a binary nonce). The zero byte between them still separates the two unambiguously, as required by NIST SP 800-108r1 (GH#896).
The coordinates of EccPoint and EccXPoint (x, y and xy), EccKey.d and the components of ElGamal keys (p, g, y and x) are now Python int objects, not Crypto.Math.Numbers.Integer. The ElGamal components can no longer be assigned.
The GMP library is no longer used, and the environment variable PYCRYPTODOME_DISABLE_GMP has no effect. Big integers are handled by the new constant-time code (see below). Crypto.Math.Numbers.Integer (an internal class) no longer supports negative values.
Elliptic curve points are decoded more strictly. Importing a NIST curve public key (SEC1, DER, PEM or OpenSSH) fails if a coordinate of the point is equal to or larger than the modulus p, as required by SEC1 (such coordinates were reduced modulo p before; EccPoint and ECC.construct() still reduce them). Ed25519 and Ed448 points (public keys, and R in signatures) are decoded as in RFC 8032: an encoding with x = 0 and the sign bit of x set, or (for Ed448) with any of the 7 low bits of the last byte set, is now rejected.
RSA keys are limited to a modulus of 16384 bits: RSA.generate(), RSA.construct() and RSA.import_key() raise ValueError for larger ones. The constant-time arithmetic is slower than GMP, and a large public key from an untrusted source could take minutes to import or use.
Crypto.Protocol.KDF.bcrypt_check() rejects (with Crypto.KdfIterationLimitError, a ValueError) a bcrypt hash whose cost is above 20, which takes about one minute: a hash from an untrusted source, which states its own cost, could otherwise tie up the CPU for a day. Use the new parameter max_cost to accept costs up to 31, as before (0 for no limit).
ChaCha20-Poly1305 with an 8-byte nonce follows RFC 8439 (the nonce is prefixed by 4 zero bytes, and the block counter has 32 bits): a single message longer than about 256 GiB is now rejected with MessageTooLongError. Before, the counter overflowed into the 4 zero bytes, which no specification defines. The ciphertext and the tag of shorter messages are unchanged.
Exporting a private key with a passphrase uses PKCS#8 with 'PBKDF2WithHMAC-SHA256AndAES256-CBC' and 600 000 iterations by default, instead of weak schemes. With RSA keys, export_key(passphrase=...) returned a PKCS#1 key with the obsolete PEM encryption (one round of MD5 and 3DES, easy to brute force), and now returns an ENCRYPTED PRIVATE KEY: the parameter pkcs of RsaKey.export_key() is now None by default (PKCS#1 for a key in the clear, as before, PKCS#8 with a passphrase). Without protection, export_key() of RSA and DSA keys and Crypto.IO.PKCS8.wrap() used 'PBKDF2WithHMAC-SHA1AndDES-EDE3-CBC', and ECC keys raised ValueError. When prot_params does not specify them, the iteration count of PBKDF2 is now 600 000 (it was 1000) and the cost N of scrypt 131 072 (it was 16 384), which meet the OWASP recommendations. The obsolete PEM encryption is only used with an explicit pkcs=1 (RSA), pkcs8=False (DSA) or use_pkcs8=False (ECC); with RSA and DSA, protection and prot_params are then rejected, instead of being silently ignored. Also, protection or prot_params without a passphrase raise ValueError, for RSA, DSA and ECC keys: before, the key was exported in the clear.
Removed Crypto.Util.number.long2str() and str2long(), aliases of long_to_bytes() and bytes_to_long() deprecated since PyCrypto.
Crypto.Hash.HMAC.new() without digestmod (which then uses MD5) emits a DeprecationWarning. The default will be removed: always pass the hash (for instance, digestmod=SHA256), as with Python's hmac module.
Crypto.Protocol.KDF.PBKDF2() without count (which then uses 1000 iterations) or without hmac_hash_module and prf (which then uses HMAC-SHA1) emits a DeprecationWarning. These weak defaults will be removed: always pass them (for instance, count=600000, hmac_hash_module=SHA256).
Crypto.Util.number.long_to_bytes() took quadratic time on large integers: now it uses int.to_bytes(), and bytes_to_long() uses int.from_bytes().
The verify() methods of MACs and AEAD modes, and bcrypt_check(), compare the tags with hmac.compare_digest(), instead of two BLAKE2s MACs with a random key.
Updated the references to superseded RFCs (PKCS#1 to RFC 8017, PKCS#5 to RFC 8018, PKCS#8 to RFC 5958), and fixed typos in error messages.
Exporting a private key with an empty passphrase of type bytes returned the key in the clear, instead of raising ValueError: for ECC keys in all formats, and for RSA (pkcs=1) and DSA (pkcs8=False) keys in DER format.
RsaKey.export_key() with format='PEM', pkcs=8, a protection but no passphrase labeled the key in the clear as ENCRYPTED PRIVATE KEY (now it raises ValueError).
Importing a private key in clear PKCS#8 format failed if a passphrase was also given; now the passphrase is ignored, as for the other formats.
Crypto.Math.Primality: the trial division by small primes in test_probable_prime() never ran (only the Miller-Rabin and Lucas tests did), which made the generation of RSA, DSA and ElGamal keys slower; and the primality tests reported 1 as a probable prime.
Key comparison: ElGamal key objects could not be compared for equality, two DSA keys always compared as equal if both were public or both private, and comparing an RSA, DSA or ElGamal key with an object of a different type raised AttributeError instead of returning False.
EdDSA signatures with S equal to the order of the curve are now rejected, as required by RFC 8032.
A few valid Ed25519 points (those where one side of the curve equation is very small, for instance with y = 3) were rejected as not on the curve.
Ed25519 and X25519: the final reduction modulo the prime (when converting a point to bytes, including the X25519 shared secret) could branch on secret data, depending on the compiler and the target (for instance, gcc -O3 on 32-bit x86, and the code without SSE2 built with clang). It is now constant time, checked with Valgrind like the rest of the elliptic curves.
The legacy can_decrypt() method of PKCS#1 ciphers raised AttributeError and not a boolean.
The update() method for SIV encryption objects now returns the cipher object, like for other AEAD modes.
Fixed the method new() of KMAC256 and TupleHash256 objects: it created a KMAC128 or TupleHash128 object instead (for KMAC, also accepting keys shorter than the 32 bytes required for KMAC256).
Added the missing documentation for Crypto.Hash.MD4.
Salsa20 never returned when encrypting or decrypting 4 GiB or more in a single call.
The package metadata declares the license as an SPDX expression (BSD-2-Clause AND Unlicense, PEP 639), which tools can process.
Crypto.Protocol.HPKE.new() raised TypeError with Python 3.9 to 3.11 when aead_id was a plain integer, and not an HPKE.AEAD member.
GH#937: Importing a malformed PEM key could take quadratic time. Thanks to Brian Willows.
Importing an encrypted OpenSSH private key (with RSA.import_key() or ECC.import_key()) had no limit on the number of rounds of bcrypt_pbkdf, which the key states: a malicious key could tie up the CPU for months. Now max_iteration_count limits it (by default, 16,384 rounds, about one minute), and a key with less than one round is rejected.
PKCS#8 keys: Crypto.PublicKey.RSA.import_key() accepted a PKCS#8 key whose private key was not an RSAPrivateKey (PKCS#1), including another, possibly encrypted, PKCS#8 key: now such a key is rejected. An encrypted PKCS#8 key with an iteration count lower than 1, or with invalid scrypt parameters (N not a power of 2 larger than 1, r or p not positive), raised other exceptions than ValueError (like ZeroDivisionError): now it is rejected with ValueError, before any key derivation.
Encrypted PKCS#8 keys protected with PBKDF2 (PBES2) were exported and imported with the slow Python implementation of PBKDF2-HMAC, and not with the C one: now they are about 50x-150x faster (SHA-1 and SHA-2).
GH#794, GH#795: the C tests failed to link on Debian armel (where the gcc sanitizers need libatomic) and mips64el (where gcc has no sanitizers); now libatomic is linked when needed, and the tests run without the sanitizers if they cannot be linked at all.
GH#320: importing an encrypted private key (PEM, PKCS#8 or OpenSSH) without a passphrase raises Crypto.PassphraseRequiredError, and with an incorrect passphrase Crypto.IncorrectPassphraseError. Both are subclasses of ValueError, like before; the messages are also clearer. This applies to import_key() of RSA, DSA and ECC, to Crypto.IO.PKCS8.unwrap() and to Crypto.IO.PEM.decode(). A new page of the documentation lists the dedicated exceptions of the library.
New exception Crypto.MessageTooLongError, raised when a cipher would process more data than allowed with the same key and nonce: CCM (it replaces CCMMessageTooLongError, still available as an alias), GCM, CTR, ChaCha20, ChaCha20-Poly1305, Salsa20, KW and KWP. It is a subclass of both ValueError and OverflowError, which those ciphers used to raise. GCM now checks the limits before processing the data, also when decrypting; and its limit on the associated data is now 264 - 1 bits (it was bytes).
All the big integer arithmetic for RSA, DSA, ElGamal and ECC runs in a new C library (Crypto.Math._nat), in constant time: the time and the memory accesses do not depend on the values of the numbers, only on their (public) size. This includes modular exponentiation and inversion, GCD, Jacobi symbol, and the Miller-Rabin and Lucas tests used to generate primes. It replaces GMP, the previous C code for modular exponentiation (IntegerCustom) and Python integers, which leak timing information; Python integers remain only as a fallback, if the C extensions are not available. The C extensions no longer depend on GMP at runtime. For now, RSA and DSA private key operations and key generation are slower than with GMP. On x86-64 CPUs with BMI2 and ADX (gcc and clang only), a build that uses those instructions is about 15-25% faster.
The NIST curves (P-192, P-224, P-256, P-384 and P-521) run on the same constant-time library (new module Crypto.PublicKey._ec_nat). Every scalar multiplication processes the full (blinded) scalar length, scans its tables in full and randomizes the coordinates, and its result is checked to be on the curve. This fixes timing leaks of the previous code, including one on the length of ECDSA nonces. It is also faster: the multiplication of the generator (signing, key generation) is up to 10x faster for P-192 and P-224, which now have precomputed tables too, and ECDH is about 1.5x-2x faster on P-192, P-224, P-256 and P-384. Multiplying the generator by a scalar of any length no longer fails.
Ed448 and X448 (Curve448) run on the same library and module, with the same countermeasures: for Ed448, a blinded scalar (by a multiple of the order of the whole group, so that it also works for points with a small-order component), signed windows with full table scans, randomized coordinates and a check of the result; for X448, the Montgomery ladder with randomized coordinates. Ed448 key generation and signing are about 10x faster (precomputed tables), verification about 2.5x, and X448 about 1.3x.
Ed25519 scalar multiplications are blinded (by a random 32-bit multiple of the order of the whole group) and randomize the coordinates. Signing is about 10% slower.
Faster signature verification. ECDSA and EdDSA verification no longer applies the side-channel countermeasures, which are useless for its public values: about 22-28% faster for P-256, P-384, P-521 and Ed448 (for the NIST curves, it also uses faster, not constant-time, formulas). Ed25519 and Ed448 points are decoded in C (RFC 8032), and copying a point no longer goes through its affine coordinates: EdDSA verification is about 2x faster for Ed25519 and 1.8x for Ed448, importing an Ed25519 public key 7x-11x faster (Ed448 2.5x-3x), and ECDSA verification on P-256 1.4x faster.
The field arithmetic of the elliptic curves has unrolled code for each curve size: scalar multiplications are 1.4x-1.7x faster for P-192, P-224 and P-256, and 1.1x-1.5x for the larger curves (more with clang than with gcc).
New function Crypto.Hash.KangarooTwelve.digest(), to hash a whole message with a single call. For messages up to 8 KiB, it is 1.3x to 3x faster than new() followed by read().
Faster hashes on x86 CPUs. With AVX2 and BMI2 (x86-64, gcc and clang only), the Keccak-based hashes use a build that takes advantage of those instructions: SHA-3, SHAKE, cSHAKE, KMAC, TupleHash and TurboSHAKE are about 1.25x-1.4x faster on long messages, and KangarooTwelve hashes 4 leaves in parallel (2x to 2.3x faster for messages of 1 MiB or more). With the SHA extensions (SHA-NI), SHA-224 and SHA-256 hash long messages about 4x faster.
New parameter threads, to run on several CPU cores (threads=0 for all of them); the output does not depend on the number of threads. It is accepted by the CTR mode (all block ciphers), the GCM and OCB modes and ChaCha20 (also XChaCha20), to encrypt and decrypt long data; by Crypto.Protocol.KDF.scrypt(), to compute the p independent flows (each thread uses its own 128 * r * N bytes of memory at the same time); and by Crypto.IO.PKCS8.wrap() and unwrap(), import_key() of RSA, DSA and ECC keys and export_key() of RSA and ECC keys, for a private key protected with scrypt (PBES2) and a parallelization parameter larger than 1.
GCM mode is about 1.5x faster on long messages, as it does not copy the data before authenticating it. The portable GHASH (for CPUs without PCLMULQDQ) no longer uses key-dependent tables: it is now constant-time, with the technique described by Thomas Pornin for BearSSL, and about 2.8x faster on 64-bit and 3.2x on 32-bit systems.
On CPython, the encrypt() and decrypt() methods of most ciphers (block cipher modes ECB, CBC, CFB, OFB, CTR, GCM and OCB, plus ChaCha20, Salsa20 and ARC4) write their result directly into the returned bytes object, without copying it (for 64 KiB or more). For long data, this is 1.3x to 1.8x faster, and up to 3x with threads. The same applies to the read() method of SHAKE128, SHAKE256, cSHAKE, TurboSHAKE and KangarooTwelve, to Crypto.Hash.KangarooTwelve.digest(), and to Crypto.Util.strxor.strxor() and strxor_c() (up to 3.6x faster).
The Salsa20 and ChaCha20 (also XChaCha20) stream ciphers are faster on long messages: about 4.8x and 4.7x on x86-64 CPUs with AVX2 and BMI2 (gcc and clang only; 8 blocks of key stream at a time), 2.6x and 2.3x with SSE2 (4 blocks at a time), and 1.3x and 1.1x otherwise. ChaCha20-Poly1305 is about 2x faster.
New parameter prot_params for DsaKey.export_key(), as for RSA and ECC keys.
GH#911: New method export() for HPKE contexts, to derive secrets (RFC 9180, Section 5.3). With the new HPKE.AEAD.EXPORT_ONLY, a context can only export secrets. Thanks to sebastian.
GH#906: New module Crypto.Cipher.oaep for RSA-OAEP encryption. It replaces Crypto.Cipher.PKCS1_OAEP. The hash function (hashmod) must always be specified, and all parameters after the key are keyword-only.
Faster key derivation functions. Crypto.Protocol.KDF.PBKDF2() is about 1.25x-1.35x faster with HMAC based on SHA-1, SHA-224, SHA-256, SHA-384 and SHA-512 (about 3.5x with HMAC-SHA256 on CPUs with SHA-NI), and about 25x-35x faster with HMAC based on SHA-3, as it now runs its inner loop in C. PBKDF1() with MD5 and SHA-1 (also used to import keys protected with the legacy PBES1 algorithms) runs its inner loop in C too: it is about 15x faster. scrypt() is about 1.15x-1.3x faster for p=1, and up to 1.7x faster for small N and large p; on Linux, its memory is backed by transparent huge pages where available. Encrypted OpenSSH private keys are imported about 1.9x faster: the two bcrypt hashes of each round of bcrypt_pbkdf run interleaved, in C.
Limits when decrypting PKCS#8 private keys (see the new page "Limits when decrypting private keys" in the documentation). New parameter max_iteration_count for import_key() of RSA, DSA and ECC keys, like the one of Crypto.IO.PKCS8.unwrap(): the maximum iteration count accepted for an encrypted PKCS#8 key. When a key is above the limit, import_key() and PKCS8.unwrap() now raise Crypto.KdfIterationLimitError (a ValueError), whose message says which value is too high and how to accept it, instead of a generic error about the key format. For scrypt, max_iteration_count now limits N·r·p/8 (that is, N for the usual r=8 and p=1), and not only N: before, a key with a large r or p could require any amount of time and memory. By default, the limit is 225. Since scrypt also does some work in proportion to r·p only, N counts as at least 1024 in the cost. New parameter max_memory for PKCS8.unwrap() and for import_key() of RSA, DSA and ECC keys: the maximum memory for the scrypt key derivation, as a total for all threads together, not for each thread (and including the memory for the input and the output of all the flows). The default is 256 MiB. Fewer threads are used if needed; if even one thread does not fit, Crypto.KdfMemoryLimitError is raised. Both are subclasses of Crypto.KdfLimitError; bcrypt_check() also raises Crypto.KdfIterationLimitError when the cost of the hash is above max_cost.
Set the environment variable PYCRYPTODOME_NO_SELFTEST when building from source to leave out the self-tests (Crypto.SelfTest), whose test keys and certificates may be reported by security scanners.
DES and Triple DES are reimplemented from scratch, in constant time (no table lookups at secret addresses), and no longer use LibTomCrypt. They are about 3.5x slower than before.
GH#908: New method seek() for Salsa20 cipher objects, to move to any position in the key stream (for instance, to start from a given block counter).
Salsa20 raises OverflowError instead of silently reusing the key stream, when the data goes beyond its end (270 bytes per nonce).
Security: the time to create an ECDSA, EdDSA or DSA signature depended on the secret nonce, enough to recover the private key with a lattice attack. D
Speed up KangarooTwelve by 50% for long messages (1MB or more).
threads for Crypto.Hash.KangarooTwelve.new(),threads=0 for all of them).max_iteration_count option. Thanks to afldl for reporting.Speed up KangarooTwelve by 50% for long messages (1MB or more).
New parameter threads for Crypto.Hash.KangarooTwelve.new(), to hash long messages on several CPU cores (threads=0 for all of them). The output does not depend on the number of threads.
GH#875: Fixed the Object Identifiers (OID) for BLAKE2.
Fixed a potential DOS attack when decrypting a password-encrypted PKCS#8 key, when the iteration count is enourmous. Decryption will now fail if it exceeds 50M iterations; the limit can be changed or removed via the new max_iteration_count option. Thanks to afldl for reporting.
Fixed the maximum number of bytes that GCM should encrypt. Thanks to Loganaden Velvindron.
Fixed ECC operations on 32-bit Windows (x86) wheels: due to a compiler bug in Visual Studio 2022, the modular inversion skipped half of the exponent bits, producing wrong results.
Fixed a bug in KangarooTwelve in the scenario where customization string is very long (8190 bytes or longer) and the message to hash is empty.
GH#922: Fixed incorrect verification of an otherwise valid ECDSA signature, when the raw x-coordinate of the computed point is not reduced. Thanks to afldl for reporting and Ville Vesilehto for the fix.
Build Windows wheel with Visual Studio 2022, from Visual Studio 2019.
Remove support for Python 3.7.
Added cipher modes Key Wrap (KW, RFC3394) and Key Wrap with Padding (KWP, RFC5649). Both are defined also in NIST SP 800-38F.
Added support for HPKE (RFC 9180).
By setting the PYCRYPTODOME_DISABLE_GMP environment variable, the GMP library will not be used even if detected.
curve to EccPoint and EccXPoint classes,Added support for TurboSHAKE128 and TurboSHAKE256.
Crypto.Hash.new() to generate a hashprot_params dictionaryFixed a side-channel leakage with OAEP decryption that could be exploited to carry out a Manger attack. Thanks to Hubert Kario.
Fixed a side-channel leakage with OAEP decryption that could be exploited to carry out a Manger attack (CVE-2023-52323). Thanks to Hubert Kario.
The update() methods of TupleHash128 and TupleHash256 objects can now hash multiple items (byte strings) at once. Thanks to Sylvain Pelissier.
update() methods of TupleHash128 and TupleHash256 objects
can now hash multiple items (byte strings) at once.
Thanks to Sylvain Pelissier.Crypto.Protocol.DH.cffi, do not use it on Windows with Python 3.12+.Added support for DER BOOLEAN encodings.
Added support for DER BOOLEAN encodings.
The library now compiles on Windows ARM64. Thanks to Niyas Sait.
Resolved issues
nonce attribute was not correctly set for XChaCha20_Poly1305 ciphers. Thanks to Liam Haber.Added support for the Counter Mode KDF defined in SP 800-108 Rev 1.
Added support for the Counter Mode KDF defined in SP 800-108 Rev 1.
Reduce the minimum tag length for the EAX cipher to 2 bytes.
An RSA object has 4 new properties for the CRT coefficients: dp, dq, invq and invq (invp is the same value as the existing u).
GH#526: improved typing for RSA.construct.
GH#534: reduced memory consumption when using a large number of cipher objects.
GH#598: fixed missing error handling for Util.number.inverse.
GH#629: improved typing for AES.new and the various mode-specific types it returns. Thanks to Greg Werbin.
GH#653: added workaround for an alleged GCC compiler bug that affected Ed25519 code compiled for AVX2.
GH#658: attribute curve of an ECC key was not always the preferred curve name, as it used to be in v3.15.0 (independently of the curve name specified when generating the key).
GH#637: fixed typing for legacy modules PKCS1_v1_5 and PKCS1_PSS, as their verify() returned a boolean.
GH#664: with OCB mode, nonces of maximum length (15 bytes) were actually used as 14 bytes nonces. After this fix, data that was encrypted in past using the (default) nonce length of 15 bytes can still be decrypted by reducing the nonce to its first 14 bytes.
GH#705: improved typing for nonce, iv, and IV parameters of cipher objects.
Build PyPy wheels only for versions 3.8 and 3.9, and not for 3.7 anymore.
Build wheels for musl Linux. Thanks to Ben Raz.
Add support for curves Ed25519 and Ed448, including export and import of keys.
Add support for curves Ed25519 and Ed448, including export and import of keys.
Add support for EdDSA signatures.
Add support for Asymmetric Key Packages (RFC5958) to import private keys.
GH#620: for Crypto.Util.number.getPrime , do not sequentially scan numbers searching for a prime.
GH#595: Fixed memory leak for GMP integers. Thanks to Witalij Siebert and Pablo Quílez.
GH#595: Fixed memory leak for GMP integers. Thanks to Witalij Siebert and Pablo Quílez.
Add support for curve NIST P-192.
Add support for curve NIST P-192.
Add support for curve NIST P-224.
Add support for curve NIST P-224.
GH#590: Fixed typing info for Crypto.PublicKey.ECC.
Relaxed ECDSA requirements for FIPS 186 signatures and accept any SHA-2 or SHA-3 hash. sign() and verify() will be performed even if the hash is stronger than the ECC key.
ECC keys in the SEC1 format can be exported and imported.
ECC keys in the SEC1 format can be exported and imported.
Add support for KMAC128, KMAC256, TupleHash128, and TupleHash256 (NIST SP-800 185).
Add support for KangarooTwelve.
GH#563: An asymmetric key could not be imported as a memoryview.
GH#566: cSHAKE128/256 generated a wrong output for customization strings longer than 255 bytes.
GH#582: CBC decryption generated the wrong plaintext when the input and the output were the same buffer. Thanks to Michael K. Ashburn.
GH#512: Especially for very small bit sizes, Crypto.Util.number.getPrime() was occasionally generating primes larger than given the bit size. Thanks t
GH#512: Especially for very small bit sizes, Crypto.Util.number.getPrime() was occasionally generating primes larger than given the bit size. Thanks to Koki Takahashi.
GH#552: Correct typing annotations for PKCS115_Cipher.decrypt().
GH#555: decrypt() method of a PKCS#1v1.5 cipher returned a bytearray instead of bytes.
GH#557: External DSA domain parameters were accepted even when the modulus (p) was not prime. This affected Crypto.PublicKey.DSA.generate() and Crypto.PublicKey.DSA.construct(). Thanks to Koki Takahashi.
Added cSHAKE128 and cSHAKE256 (of SHA-3 family). Thanks to Michael Schaffner.
GH#558: The flag RTLD_DEEPBIND passed to dlopen() is not well supported by address sanitizers. It is now possible to set the environment variable PYCRYPTDOME_DISABLE_DEEPBIND to drop that flag and allow security testing.
Output of Crypto.Util.number.long_to_bytes() was not always a multiple of blocksize.
Output of Crypto.Util.number.long_to_bytes() was not always a multiple of blocksize.
GH#376: Fixed symbol conflict between different versions of libgmp.
GH#376: Fixed symbol conflict between different versions of libgmp.
GH#481: Improved robustness of PKCS#1v1.5 decryption against timing attacks.
GH#506 and GH#509: Fixed segmentation faults on Apple M1 and other Aarch64 SoCs, when the GMP library was accessed via ctypes. Do not use GMP's own sscanf and snprintf routines: instead, use simpler conversion routines.
GH#510: Workaround for cffi calling ctypes.util.find_library(), which invokes gcc and ld on Linux, considerably slowing down all imports. On certain configurations, that may also leave temporary files behind.
GH#517: Fix RSAES-OAEP, as it didn't always fail when zero padding was incorrect.
Added support for SHA-3 hash functions to HMAC.
The Windows wheels of Python 2.7 now require the VS2015 runtime to be installed in the system, because Microsoft stopped distributing the VS2008 compiler in April 2021. VS2008 was used to compile the Python 2.7 extensions.
Python 3 wheels use abi3 ABI tag.
Python 3 wheels use abi3 ABI tag.
Remove Appveyor CI.
GH#435: Fixed Crypto.Util.number.size for negative numbers.
GH#435: Fixed Crypto.Util.number.size for negative numbers.
Build Python 3.9 wheels on Windows.
GH#426: The Shamir's secret sharing implementation is not actually compatible with ssss. Added an optional parameter to enable interoperability.
GH#426: The Shamir's secret sharing implementation is not actually compatible with ssss. Added an optional parameter to enable interoperability.
GH#427: Skip altogether loading of gmp.dll on Windows.
GH#420: Fix incorrect CFB decryption when the input and the output are the same buffer.
Speed up Shamir's secret sharing routines. Thanks to ncarve.
GH#381: Make notarization possible again on OS X when using wheels. Thanks to Colin Atkinson.
GH#381: Make notarization possible again on OS X when using wheels. Thanks to Colin Atkinson.
Fix building of wheels for OS X by explicitly setting sysroot location.
Fix building of wheels for OS X by explicitly setting sysroot location.
GH#341: Prevent key_to_english from creating invalid data when fed with keys of length not multiple of 8. Thanks to vstoykovbg.
GH#341: Prevent key_to_english from creating invalid data when fed with keys of length not multiple of 8. Thanks to vstoykovbg.
GH#347: Fix blocking RSA signing/decryption when key has very small factor. Thanks to Martijn Pieters.
GH#308: Align stack of functions using SSE2 intrinsics to avoid crashes, when compiled with gcc on 32-bit x86 platforms.
GH#308: Align stack of functions using SSE2 intrinsics to avoid crashes, when compiled with gcc on 32-bit x86 platforms.
GH#308: Avoid allocating arrays of __m128i on the stack, to cope with buggy compilers.
Add Python 3.8 wheels for Mac.
GH#308: Avoid allocating arrays of __m128i on the stack, to cope with buggy compilers.
GH#322: Remove blanket -O3 optimization for gcc and clang, to cope with buggy compilers.
GH#337: Fix typing stubs for signatures.
GH#338: Deal with gcc installations that don't have x86intrin.h.
Add Python 3.8 wheels for Linux and Windows.
Add Python 3.8 wheels for Linux and Windows.
GH#328: minor speed-up when importing RSA.
Add support for loading PEM files encrypted with AES256-CBC.
Add support for loading PEM files encrypted with AES256-CBC.
Add support for XChaCha20 and XChaCha20-Poly1305 ciphers.
Add support for bcrypt key derivation function (Crypto.Protocol.KDF.bcrypt).
Add support for left multiplication of an EC point by a scalar.
Add support for importing ECC and RSA keys in the new OpenSSH format.
GH#312: it was not possible to invert an EC point anymore.
GH#316: fix printing of DSA keys.
GH#317: DSA.generate() was not always using the randfunc input.
GH#285: the MD2 hash had block size of 64 bytes instead of 16; as result the HMAC construction gave incorrect results.
GH#291: fix strict aliasing problem, emerged with GCC 9.1.
GH#291: fix strict aliasing problem, emerged with GCC 9.1.
Add support for loading PEM files encrypted with AES192-CBC and AES256-GCM.
Add support for loading PEM files encrypted with AES192-CBC and AES256-GCM.
When importing ECC keys in PEM format, ignore the redundant EC PARAMS section that was included by certain openssl commands.
repr() did not work for ECC.EccKey objects.
Fix installation in development mode (setup install develop or pip install -e .).
Minimal length for Blowfish cipher is 32 bits, not 40 bits.
Various updates to docs.
Speed-up ECC performance. ECDSA is 33 times faster on the NIST P-256 curve.
Speed-up ECC performance. ECDSA is 33 times faster on the NIST P-256 curve.
Added support for NIST P-384 and P-521 curves.
EccKey has new methods size_in_bits() and size_in_bytes().
Support HMAC-SHA224, HMAC-SHA256, HMAC-SHA384, and HMAC-SHA512 in PBE2/PBKDF2.
DER objects were not rejected if their length field had a leading zero.
Allow legacy RC2 ciphers to have 40-bit keys.
ASN.1 Object IDs did not allow the value 0 in the path.
point_at_infinity() becomes an instance method for Crypto.PublicKey.ECC.EccKey, from a static one.
GH#258: False positive on PSS signatures when externally provided salt is too long.
GH#258: False positive on PSS signatures when externally provided salt is too long.
Include type stub files for Crypto.IO and Crypto.Util.
GH#242: Fixed compilation problem on ARM platforms.
GH#242: Fixed compilation problem on ARM platforms.
Added type stubs to enable static type checking with mypy. Thanks to Michael Nix.
Added type stubs to enable static type checking with mypy. Thanks to Michael Nix.
New update_after_digest flag for CMAC.
GH#232: Fixed problem with gcc 4.x when compiling ghash_clmul.c.
GH#238: Incorrect digest value produced by CMAC after cloning the object.
Method update() of an EAX cipher object was returning the underlying CMAC object, instead of the EAX object itself.
Method update() of a CMAC object was not throwing an exception after the digest was computed (with digest() or verify()).
Added support for Poly1305 MAC (with AES and ChaCha20 ciphers for key derivation).
Added support for Poly1305 MAC (with AES and ChaCha20 ciphers for key derivation).
Added support for ChaCha20-Poly1305 AEAD cipher.
New parameter output for Crypto.Util.strxor.strxor, Crypto.Util.strxor.strxor_c, encrypt and decrypt methods in symmetric ciphers (Crypto.Cipher package). output is a pre-allocated buffer (a bytearray or a writeable memoryview) where the result must be stored. This requires less memory for very large payloads; it is also more efficient when encrypting (or decrypting) several small payloads.
GH#266: AES-GCM hangs when processing more than 4GB at a time on x86 with PCLMULQDQ instruction.
Drop support for Python 3.3.
Remove Crypto.Util.py3compat.unhexlify and Crypto.Util.py3compat.hexlify.
With the old Python 2.6, use only ctypes (and not cffi) to interface to native code.
GH#198: Fix vulnerability on AESNI ECB with payloads smaller than 16 bytes (CVE-2018-15560).
GH#198: Fix vulnerability on AESNI ECB with payloads smaller than 16 bytes (CVE-2018-15560).
Fixed ResouceWarnings and DeprecationWarnings.
GH#187: Fixed incorrect AES encryption/decryption with AES acceleration on x86 due to gcc's optimization and strict aliasing rules.
GH#188: More prime number candidates than necessary where discarded as composite due to the limited way D values were searched in the Lucas test.
Fixed ResouceWarnings and DeprecationWarnings.
Workaround for Python 3.7.0 bug on Windows (https://bugs.python.org/issue34108).
Build Python 3.7 wheels on Linux, Windows and Mac.
Build Python 3.7 wheels on Linux, Windows and Mac.
GH#178: Rename _cpuid module to make upgrades more robust.
More meaningful exceptions in case of mismatch in IV length (CBC/OFB/CFB modes).
Fix compilation issues on Solaris 10/11.
GH#175: Fixed incorrect results for CTR encryption/decryption with more than 8 blocks.
GH#175: Fixed incorrect results for CTR encryption/decryption with more than 8 blocks.
Added Google Wycheproof tests (https://github.com/google/wycheproof) for RSA, DSA, ECDSA, GCM, SIV, EAX, CMAC.
Added Google Wycheproof tests (https://github.com/google/wycheproof) for RSA, DSA, ECDSA, GCM, SIV, EAX, CMAC.
New parameter mac_len (length of MAC tag) for CMAC.
In certain circumstances (at counter wrapping, which happens on average after 32 GB) AES GCM produced wrong ciphertexts.
Method encrypt() of AES SIV cipher could be still called, whereas only encrypt_and_digest() is allowed.
Introduced export_key and deprecated exportKey for DSA and RSA key objects.
Introduced export_key and deprecated exportKey for DSA and RSA key objects.
Ciphers and hash functions accept memoryview objects in input.
Added support for SHA-512/224 and SHA-512/256.
Reintroduced Crypto.__version__ variable as in PyCrypto.
Fixed compilation problem with MinGW.
GH#142. Fix mismatch with declaration and definition of addmul128.
GH#142. Fix mismatch with declaration and definition of addmul128.
Import and export of ECC curves in compressed form.
Import and export of ECC curves in compressed form.
The initial counter for a cipher in CTR mode can be a byte string (in addition to an integer).
Faster PBKDF2 for HMAC-based PRFs (at least 20x for short passwords, more for longer passwords). Thanks to Christian Heimes for pointing out the implementation was under-optimized.
The salt for PBKDF2 can be either a string or bytes (GH#67).
Ciphers and hash functions accept data as bytearray, not just binary strings.
The old SHA-1 and MD5 hash functions are available even when Python's own hashlib does not include them.
Without libgmp, modular exponentiation (since v3.4.8) crashed on 32-bit big-endian systems.
Removed support for Python < 2.6.
GH#121. the record list was still not correct due to PEP3147 and __pycache__ directories. Thanks again to John O'Brien.
GH#121. the record list was still not correct due to PEP3147 and __pycache__ directories. Thanks again to John O'Brien.
More meaningful error messages while importing an ECC key.
More meaningful error messages while importing an ECC key.
GH#123 and #125. The SSE2 command line switch was not always passed on 32-bit x86 platforms.
GH#121. The record list (--record) was not always correctly filled for the pycryptodomex package. Thanks to John W. O'Brien.
Added a native extension in pure C for modular exponentiation, optimized for SSE2 on x86. In the process, we drop support for the arbitrary arithmetic
Added a native extension in pure C for modular exponentiation, optimized for SSE2 on x86. In the process, we drop support for the arbitrary arithmetic library MPIR on Windows, which is painful to compile and deploy. The custom modular exponentiation is 130% (160%) slower on an Intel CPU in 32-bit (64-bit) mode, compared to MPIR. Still, that is much faster that CPython's own pow() function which is 900% (855%) slower than MPIR. Support for the GMP library on Unix remains.
Added support for manylinux wheels.
Support for Python 3.7.
The DSA parameter 'p' prime was created with 255 bits cleared (but still with the correct strength).
GH#106. Not all docs were included in the tar ball. Thanks to Christopher Hoskin.
GH#109. ECDSA verification failed for DER encoded signatures. Thanks to Alastair Houghton.
Human-friendly messages for padding errors with ECB and CBC.
API documentation is made with sphinx instead of epydoc.
API documentation is made with sphinx instead of epydoc.
Start using importlib instead of imp where available.
GH#82. Fixed PEM header for RSA/DSA public keys.
GH#65. Keccak, SHA3, SHAKE and the seek functionality for ChaCha20 were not working on big endian machines. Fixed. Thanks to Mike Gilbert.
GH#65. Keccak, SHA3, SHAKE and the seek functionality for ChaCha20 were not working on big endian machines. Fixed. Thanks to Mike Gilbert.
A few fixes in the documentation.
The library can also be compiled using MinGW.
The library can also be compiled using MinGW.
[Security] Removed implementation of deprecated "quick check" feature of PGP block cipher mode.
Removed use of alloca().
[Security] Removed implementation of deprecated "quick check" feature of PGP block cipher mode.
Improved the performance of scrypt by converting some Python to C.
Undefined warning was raised with libgmp version < 5
Undefined warning was raised with libgmp version < 5
Forgot inclusion of alloca.h
Fixed a warning about type mismatch raised by recent versions of cffi
One column per quarter.