cryptocomrade: Cryptography in Golang

In this blog post I will share with you how I write cryptography software in Go. This is a work in progress. I’ll periodically update this so as to fix mistakes and add more content. The cryptocomrade git repo is meant to accompany this blog post by providing code samples with a creative commons license (CC0):

However I only added a few simple hash examples so far. I plan to add more soon.

Cryptography Application Best Practices

Cryptography applications should try to prevent leaking timing information. Performing constant time comparisons via the “subtle” module can help with this goal:

Use the “crypto/subtle” package:

Here’s an example of the constant time byte slice comparison:

    if subtle.ConstantTimeCompare(x, y) == 0 {
        return errors.New("values not equal")

And here’s an example in the wild written by Yawning Angel:

	if subtle.ConstantTimeCompare(pkt[macOff:macOff+crypto.MACLength], mac) != 1 {
		return nil, replayTag[:], nil, errors.New("sphinx: invalid packet, MAC mismatch")

In this case the author Yawning Angel is using it to compare two MACs, a good use case for the “crypto/subtle” module and this is exactly why the “crypto/hmac” module has a wrapper function for “subtle.ConstantTimeCompare” called “Equal” which returns a bool. Use either one but pick one and use it to avoid leaking timing information.

Cryptography applications should also try to prevent the capture of cryptographic key material by various attacks including the cold boot attacks. It’s best to not leave key material lying around in the system memory unnecessarily. To that end Memguard may be of use.

MemGuard: Software enclave for storage of sensitive information in memory.

See for the complete feature set.

In the author’s words:

This package attempts to reduce the likelihood of sensitive data being
exposed when in memory. It aims to support all major operating systems
and is written in pure Go.

Memguard has thorough documentation and plenty of examples, however, here’s an extensive example found in the wild:

It’s certainly possible to use Memguard and not get any benefit from it. This can happen if you are making superfluous copies of your key material. Such copies essentially defeat the purpose of using Memguard. Instead you are suppose to pass around references to Memguard’s opaque types and use those to guard access to the key material.

If you are filling up a buffer with random bytes you should do so using “crypto/rand” and NOT “math/rand”. If this operation fails just panic because if the device ran out of entropy no cryptography can be safely used.

That having been said, sometimes when we write crypto code we choose to ignore errors where appropriate to do so like in the case where the io.Read interface semantics don’t exactly line up with the API being used. What I mean is there are cases where a cryptography API like a stream cipher will ALWAYS return successfully from an io.Read operation. Therefore to ignore io.Read errors in such cases is acceptable and almost always makes your API cleaner because your function signatures will be shorter when not having to return an error.

I suggest reading these two sources for general implementation guidelines, although they are not specific to Go:

Cryptographic Hashing

Use Blake2 because it’s harder, better, faster, stronger.

The Blake2 website says:

BLAKE2b is more efficient on 64-bit CPUs and BLAKE2s is more efficient on
8-bit, 16-bit, or 32-bit CPUs

So it seems to me that Blake2b is good enough for most use cases since many mobile devices these days have 64bit CPUs. Blake2 is very versitile and has multiple hash output lengths, can be used as a KDF with it’s XOF (extensible output function) or a MAC (message authentication code) with it’s built-in keying mechanism. Blake2 is not vulnerable to the length extension attack so an HMAC construction is not needed.

The rfc7693 states that:

Both BLAKE2b and BLAKE2s are believed to be highly secure and perform
well on any platform, software, or hardware. BLAKE2 does not require
a special "HMAC" (Hashed Message Authentication Code) construction
for keyed message authentication as it has a built-in keying

I suggest using the Blake2b implementation from the supplementary Go cryptography libraries:

Hash a byte slice:

// SimplexHash is the simplest usage of a hash function
// and is suitable if you only have one use case.
func SimplexHash(data []byte) []byte {
	out := blake2b.Sum256(data)
	return out[:]

Instead you probably want to use something like this for multiple use cases:

// Hash hashes the data with contextInfo used as the key
// for blake2's keying mechanism. contextInfo is meant to
// be a human readable string that ensures each use case of
// the hash function will have different outputs.
func Hash(contextInfo string, data []byte) ([]byte, error) {
	if len(contextInfo) > blake2b.Size {
		return nil, errors.New("contextInfo size error")
	h, err := blake2b.New256([]byte(contextInfo))
	if err != nil {
		return nil, err
	_, err = h.Write(data)
	if err != nil {
		return nil, err
	return h.Sum(nil), nil

Key Derivation Function

Just use HKDF. Specifically, HKDF-SHA256 is probably sufficient for most use cases when your initial key material is 32 bytes in length. You MUST only feed your KDF a key which has uniform entropy. If you need to derive keys from a password or passphrase then you should first use a “key stretching function” which I discuss in the Hashing Passwords section below.

Remember how I mentioned above that Blake2 is versatile and can be used as a KDF and a MAC? Well, I suggest we NOT using Blake2 for all of these use cases in a single application because if you do and it becomes weakened or compromised by a cryptographic attack then several parts of your construction will be broken instead of just one.

(But if you insist on using Blake2b as a KDF this is how you’d do it: here’s an example: blah blah url goes here)

I suggest using the HKDF implementation from the supplementary Go cryptography libraries:

Here’s a simple example usage:

// KDF returns a slice of derive keysNum number of derived keys of size keySize.
// The given secret must be a uniform entropy secret (ie not a password) and
// info is an optional non-secret which may be omitted.
// Note: in practice you probably don't want to use this particular function
// but instead use the HKDF directly.
func KDF(secret, salt, info []byte, keysNum, keySize int) ([][]byte, error) {
	if len(salt) != hash().Size() {
		return nil, errors.New("wrong salt size")
	hash := sha256.New
	hkdf := hkdf.New(hash, secret, salt, info)
	var keys [][]byte
	for i := 0; i < keysNum; i++ {
		key := make([]byte, keySize)
		if _, err := io.ReadFull(hkdf, key); err != nil {
		keys = append(keys, key)
	return keys, nil

The hkdf type implements the io.Reader interface so you can read from it with any size byte slice that you’d like. Usually I use it to generate 32 byte keys.

Hashing Passwords

Use Argon2 because it won the Password Hashing Competition. It’s suitable to use if you are storing password hashes in a database or if you have to derive cryptographic keys from a password or passphrase.

I suggest using the Argon2id implementation from the supplementary Go cryptography libraries:

Argon2id is side-channel resistant and provides better brute-
force cost savings due to time-memory tradeoffs than Argon2i. 

This example consumes 64MB of memory. So depending on your device and your application you’ll want to adjust this parameter:

// HashPassword returns a 32 byte cryptographic key given
// a password and an entropic salt.
func HashPassword(password, salt []byte) []byte {
	return argon2.IDKey(password, salt, 1, 64*1024, 4, 32)

Note that the last parameter is key size, thus you won’t need to use Argon2 with HKDF because you can specify any key size you’d like. However the previously mentioned HKDF uses the io.Reader interface which makes it quite easy to derive multiple keys without having to fiddle with byte slices. It wouldn’t hurt to process the output of Argon2 with HKDF and the performance hit is probably negligible.

Message Authentication Code

First choice: Just use NaCl auth because it exists and it’s easy to use. NaCl auth uses HMAC-SHA512:

Otherwise use Blake2b’s keyed hash feature for a MAC:

// Blake2bMAC is a MAC that uses Blake2b's keyed hash mechanism
// instead of an HMAC construction. The output is of size MACSize.
func Blake2bMAC(key, data []byte) []byte {
	h, err := blake2b.New(MACSize, key)
	if err != nil {
	return h.Sum(nil)

// ValidBlake2bMAC reports whether Blake2b messageMAC is a valid MAC tag for message.
func ValidBlake2bMAC(message, messageMAC, key []byte) bool {
	expectedMAC := Blake2bMAC(key, message)
	return hmac.Equal(messageMAC, expectedMAC)

If you need the MAC tag to be 32 bytes long are already using Blake2b as a hash then you should select a different MAC such as HMAC-SHA256.

package mac

import (

// HMACSHA256 returns the HMAC-SHA256 authentication code for the given message and key.
func HMACSHA256(message, key []byte) []byte {
	mac := hmac.New(sha256.New, key)
	return mac.Sum(nil)

// ValidMAC reports whether messageMAC is a valid HMAC tag for message.
func ValidHMACSHA256(message, messageMAC, key []byte) bool {
	return hmac.Equal(messageMAC, HMACSHA256(message, key))

AEAD Ciphers

AEAD stands for Authenticated Encryption with Associated Data. Unless the AEAD has a synthetic IV, nonce reuse is not safe. The way this situation was handled in the past was to turn the nonce into a counter and always increment the nonce with each use. This turns out to be very difficult to do properly in practice. So the preferred solution is instead to use bigger nonces that can safely be randomly generated. That is, the nonce size should be big enough that randomly generating it has an extremely low probability of a nonce being reused.

But I would instead recommend the AEAD known as XSalsa20Poly1305 or XChacha20Poly1305 which is composed of the Poly1305 MAC and the XSalsa20 or the XChacha20 stream cipher. Salsa20 was the original but with it’s 8 byte nonce it required careful nonce management. The X variant, XSalsa20 increased the nonce size to 24 bytes. According to the libsodium docs:

But with XSalsa20's longer nonce, it is safe to generate nonces using
randombytes_buf() for every message encrypted with the same key without
having to worry about a collision.

I’m not clear on what the differences are between XSalsa20 and XChacha20 except that initially Salsa20 had a security proof and Chacha20 did not because of being newer. Whereas they now both have security proofs and so do their bigger nonce variants.

The older recommendation would be to use the AES-GCM AEAD but I think we’re past that phase in cryptography history now (although NIST lags behind of course). In agl’s words:

These are primitives developed by Dan Bernstein and are fast,
secure, have high quality, public domain implementations, are
naturally constant time and have nearly perfect key agility.

Use NaCl secretbox to symmetrically encrypt small messages (small enough to fit in memory). NaCl secretbox uses XSalsa20 and Poly1305. I really like the NaCl secretbox API for golang, it’s very simple:

In the following example I randomly generate a nonce for each message which is to be encrypted. The ciphertext returned has got the nonce in the beginning 24 bytes so that it can be used for the subsequent decryption. For example:

package cryptocomrade

import (


const (
	nonceSize = 24

var (
	errOpen = errors.New("failed to open envelope")

// SecretKey encapsulates a secret key so that it can be used
// in AEAD cipher operations methods: Open and Seal.
type SecretKey struct {
	key *[32]byte

// Load loads a SecretKey.
func Load(b *[32]byte) *SecretKey {
	return &SecretKey{
		key: b,

// New returns a new SecretKey.
func New() *SecretKey {
	b := [32]byte{}
	_, err := rand.Read(b[:])
	if err != nil {
	return &SecretKey{
		key: &b,

// Bytes returns the secret key as a byte slice.
func (s *SecretKey) Bytes() []byte {
	return s.key[:]

// Open decrypts the ciphertext and returns the plaintext or an error.
func (s *SecretKey) Open(ciphertext []byte) ([]byte, error) {
	nonce := [24]byte{}
	copy(nonce[:], ciphertext[:nonceSize])
	out, ok := secretbox.Open(nil, ciphertext[nonceSize:], &nonce, s.key)
	if !ok {
		return nil, errOpen
	return out, nil

// Seal encrypts the plaintext and returns the ciphertext.
func (s *SecretKey) Seal(plaintext []byte) []byte {
	nonce := [24]byte{}
	_, err := rand.Read(nonce[:])
	if err != nil {
	out := make([]byte, nonceSize)
	copy(out, nonce[:])
	ciphertext := secretbox.Seal(out, plaintext, &nonce, s.key)
	return ciphertext

Another good choice would be to use XChacha20Poly1305 AEAD cipher directly:

If your encryption payload is too big to fit in memory then you could use the “age” API. See the next section “Encrypting Files”.

Asymmetric Authenticated Encryption

Use NaCl box box to asymmetrically encrypt small messages. Box uses Curve25519, XSalsa20 and Poly1305.

Encrypting Files

This section is also known as the Encrypting Bigass Files section.

If you have some bigass files to encrypt then using an AEAD cipher directly without a buffering layer will result in loading the entire plaintext into memory all at once. Not good if your file is bigger than your device’s available memory. Instead just use age because it exists and is easy to use. Although age is a commandline tool it’s also a cleanly written Go encryption library.

Age has support for encrypting using symmetric and asymmetric ciphers.


Source code:

Golang API Docs:

Cryptographic Transports

I cannot emphasize enough how great the Noise Protocol Framework is. It really is the best choice for a cryptographic transport. Instead of negiotiating a cipher suite at runtime, Noise makes the developer choose all the protocol design criteria during the development phase which is why it’s called a “protocol framework”.

In general I like to recommend everyone use the “XX” handshake pattern with curve25519, ChaChaPoly AEAD cipher and the Blake2b hash. This implies using the Noise protocol descriptor string of: Noise_XX_25519_ChaChaPoly_Blake2b

That having been said, there are those who would sacrifice security for convenience and use a different Noise handshake pattern because of such features as zero round trip time encryption. This means that they get to send an encrypted message on the first handshake message whereas XX is a threeway handshake.

All that having been said, there’s not usually a good reason to not use a Noise protocol and incorporates a post quantum key encapsulation mechanism.

Here’s an example of Noise being used with the Kyber post quantum key encapsulation mechanism with the Noise hybrid forward secret protocol modifier, it’s full handshake protocol descriptor string is:


Here’s the XXhfs handshake pattern:

-> e, e1
<- e, ee, ekem1, s, es
-> s, se

TODO: explain what this means

HFS: Hybrid forward secrecy in the case of Noise means that the handshake encapsulates the key exchange using Curve25519 AND Kyber both. That way if one of them is “broken” then we still have the security guarantees of the handshake. Furthermore using a PQ KEM like Kyber protects recorded ciphertext from future attacks by an adversary with a sufficiently powerful quantum computer.

The example code:

To use Kyber with Noise in Go you can use the Katzenpost fork of Noise, here:

But beware, this isn’t the typical advice most people give about using Noise. It’s very possible that if you use a Noise HFS Kyber protocol that some cryptography people may mock you for being too paranoid. ;-)

TODO: write more about Noise here

Summary Cryptographic Primitive Recommendations


In this blog post I haven’t mentioned TLS but you can read about how to use TLS in Go here: