paulmillr/noble-hashes

noble-hashes

Audited & minimal JS implementation of hash functions, MACs and KDFs.

• 🔒 Audited by an independent security firm
• 🪶 Minimal: 2.7KB (gzipped) sha256, unused code is excluded from your builds
• 🏎 Fast: hand-optimized for caveats of JS engines
• 🔍 Reliable: chained / ACVP tests ensure correctness
• 🔁 No unrolled loops: makes it easier to verify and reduces source code size up to 5x
• 🦘 Includes SHA, RIPEMD, BLAKE, HMAC, HKDF, PBKDF, Scrypt, Argon2
• 🥈 Wrapper with identical API over native WebCrypto

The library's initial development was funded by Ethereum Foundation.

This library belongs to noble cryptography

noble cryptography — high-security, easily auditable set of contained cryptographic libraries and tools.

• Zero or minimal dependencies
• Highly readable TypeScript / JS code
• PGP-signed releases and transparent NPM builds
• All libraries: ciphers, curves, hashes, post-quantum, 5kb secp256k1 / ed25519
• WASM version: awasm-noble
• Check out the homepage for reading resources, documentation, and apps built with noble

Usage

npm install @noble/hashes

deno add jsr:@noble/hashes

We support all major platforms and runtimes. For React Native, you may need a polyfill for getRandomValues. A standalone file noble-hashes.js is also available.

// import * from '@noble/hashes'; // Error: use sub-imports, to ensure small app size
import { sha256 as noble_sha256 } from '@noble/hashes/sha2.js';
const hash = noble_sha256(Uint8Array.from([0xca, 0xfe, 0x01, 0x23]));

// Available modules
import { sha256, sha384, sha512, sha224, sha512_224, sha512_256 } from '@noble/hashes/sha2.js';
import {
  sha3_256, sha3_512,
  keccak_256, keccak_512,
  shake128, shake256,
} from '@noble/hashes/sha3.js';
import {
  cshake256, turboshake256, kmac256, tuplehash256,
  kt128, kt256, keccakprg,
} from '@noble/hashes/sha3-addons.js';
import { blake3 } from '@noble/hashes/blake3.js';
import { blake2b, blake2s } from '@noble/hashes/blake2.js';
import { blake256, blake512 } from '@noble/hashes/blake1.js';
import { sha1, md5, ripemd160 } from '@noble/hashes/legacy.js';
import { hmac } from '@noble/hashes/hmac.js';
import { hkdf } from '@noble/hashes/hkdf.js';
import { pbkdf2, pbkdf2Async } from '@noble/hashes/pbkdf2.js';
import { scrypt, scryptAsync } from '@noble/hashes/scrypt.js';
import { argon2d, argon2i, argon2id } from '@noble/hashes/argon2.js';
import { eskdf } from '@noble/hashes/eskdf.js';
import * as webcrypto from '@noble/hashes/webcrypto.js';
// const { sha256, sha384, sha512, hmac, hkdf, pbkdf2 } = webcrypto;
import * as utils from '@noble/hashes/utils.js';
const { bytesToHex, concatBytes, equalBytes, hexToBytes } = utils;

• sha2: sha256, sha384, sha512
• sha3: FIPS, SHAKE, Keccak
• sha3-addons: cSHAKE, KMAC, KT128, TurboSHAKE
• blake1, blake2, blake3
• legacy: sha1, md5, ripemd160
• MACs: hmac | kmac | blake3 key mode
• KDFs: hkdf | pbkdf2 | scrypt | argon2 | eskdf
• webcrypto: friendly wrapper
• utils
• Security | Speed | Contributing & testing | License

Implementations

Hash functions:

• sha256(): receive & return Uint8Array
• sha256.create().update(a).update(b).digest(): support partial updates
• blake3.create({ context: 'e', dkLen: 32 }): can have options
• support little-endian architecture; also experimentally big-endian
• can hash up to 4GB per chunk, with any amount of chunks

sha2: sha256, sha384, sha512 and others

import { sha224, sha256, sha384, sha512, sha512_224, sha512_256 } from '@noble/hashes/sha2.js';
const res = sha256(Uint8Array.from([0xbc])); // basic
for (let hash of [sha256, sha384, sha512, sha224, sha512_224, sha512_256]) {
  const arr = Uint8Array.from([0x10, 0x20, 0x30]);
  const a = hash(arr);
  const b = hash.create().update(arr).digest();
}

Check out RFC 6234 and the paper on truncated SHA512/256.

sha3: FIPS, SHAKE, Keccak

import {
  sha3_224, sha3_256, sha3_384, sha3_512,
  keccak_224, keccak_256, keccak_384, keccak_512,
  shake128, shake256,
} from '@noble/hashes/sha3.js';
for (let hash of [
  sha3_224, sha3_256, sha3_384, sha3_512,
  keccak_224, keccak_256, keccak_384, keccak_512,
]) {
  const arr = Uint8Array.from([0x10, 0x20, 0x30]);
  const a = hash(arr);
  const b = hash.create().update(arr).digest();
}
const shka = shake128(Uint8Array.from([0x10]), { dkLen: 512 });
const shkb = shake256(Uint8Array.from([0x30]), { dkLen: 512 });

Check out FIPS-202, Website.

Check out the differences between SHA-3 and Keccak

sha3-addons: cSHAKE, KMAC, KT128, TurboSHAKE

import {
  cshake128, cshake256, kt128, kt256,
  keccakprg, kmac128, kmac256,
  parallelhash256, tuplehash256,
  turboshake128, turboshake256,
} from '@noble/hashes/sha3-addons.js';
const data = Uint8Array.from([0x10, 0x20, 0x30]);
const personalization = new TextEncoder().encode('def');
const ec1 = cshake128(data, { personalization });
const ec2 = cshake256(data, { personalization });
const et1 = turboshake128(data);
const et2 = turboshake256(data, { D: 0x05 });
// tuplehash(['ab', 'c']) !== tuplehash(['a', 'bc']) !== tuplehash([data])
const et3 = tuplehash256([new TextEncoder().encode('ab'), new TextEncoder().encode('c')]);
// Not parallel in JS (similar to blake3 / kt128), added for compat
const ep1 = parallelhash256(data, { blockLen: 8 });
const kk = Uint8Array.from([0xca]);
const ek10 = kmac128(kk, data);
const ek11 = kmac256(kk, data);
const ek12 = kt128(data); // kangarootwelve 128-bit
const ek13 = kt256(data); // kangarootwelve 256-bit
// pseudo-random generator, first argument is capacity. XKCP recommends 254 bits capacity for 128-bit security strength.
const p = keccakprg(254);
p.addEntropy(Uint8Array.from([1, 2, 3]));
const rand1b = p.randomBytes(32);

• cSHAKE, KMAC, TupleHash, ParallelHash + XOF are available, matching NIST SP 800-185
• Reduced-round Keccak KT128 (KangarooTwelve 🦘, K12) and TurboSHAKE are available, matching RFC 9861.
• KeccakPRG: pseudo-random generator based on Keccak

blake1, blake2, blake3

import { blake224, blake256, blake384, blake512 } from '@noble/hashes/blake1.js';
import { blake2b, blake2s } from '@noble/hashes/blake2.js';
import { blake3 } from '@noble/hashes/blake3.js';

for (let hash of [blake224, blake256, blake384, blake512, blake2b, blake2s, blake3]) {
  const arr = Uint8Array.from([0x10, 0x20, 0x30]);
  const a = hash(arr);
  const b = hash.create().update(arr).digest();
}

// blake2 advanced usage
const ab = Uint8Array.from([0x01]);
const txt = new TextEncoder();
blake2s(ab);
blake2s(ab, { key: new Uint8Array(32) });
blake2s(ab, { personalization: txt.encode('pers1234') });
blake2s(ab, { salt: txt.encode('salt1234') });
blake2b(ab);
blake2b(ab, { key: new Uint8Array(64) });
blake2b(ab, { personalization: txt.encode('pers1234pers1234') });
blake2b(ab, { salt: txt.encode('salt1234salt1234') });

// blake3 advanced usage
blake3(ab);
blake3(ab, { dkLen: 256 });
blake3(ab, { key: new Uint8Array(32) });
blake3(ab, { context: txt.encode('application-name') });

• Blake1 is legacy hash, one of SHA3 proposals. It is rarely used anywhere. See pdf.
• Blake2 is popular fast hash. blake2b focuses on 64-bit platforms while blake2s is for 8-bit to 32-bit ones. See RFC 7693, Website
• Blake3 is faster, reduced-round blake2. See Website & specs

legacy: sha1, md5, ripemd160

SHA1 (RFC 3174), MD5 (RFC 1321) and RIPEMD160 (ISO/IEC 10118-3) legacy, weak hash functions. Don't use them in a new protocol. What "weak" means:

• Collisions can be made with 2^24 effort in MD5 (seconds on commodity hardware), 2^61 in SHA1 (demonstrated in practice), 2^80 in RIPEMD160.
• No practical pre-image attacks (only theoretical, 2^123.4)
• HMAC seems kinda ok: https://datatracker.ietf.org/doc/html/rfc6151

import { md5, ripemd160, sha1 } from '@noble/hashes/legacy.js';
for (let hash of [md5, ripemd160, sha1]) {
  const arr = Uint8Array.from([0x10, 0x20, 0x30]);
  const a = hash(arr);
  const b = hash.create().update(arr).digest();
}

hmac

import { hmac } from '@noble/hashes/hmac.js';
import { sha256 } from '@noble/hashes/sha2.js';
const key = new Uint8Array(32).fill(1);
const msg = new Uint8Array(32).fill(2);
const mac1 = hmac(sha256, key, msg);
const mac2 = hmac.create(sha256, key).update(msg).digest();

Conforms to RFC 2104.

hkdf

import { hkdf } from '@noble/hashes/hkdf.js';
import { randomBytes } from '@noble/hashes/utils.js';
import { sha256 } from '@noble/hashes/sha2.js';
const inputKey = randomBytes(32);
const salt = randomBytes(32);
const info = new TextEncoder().encode('application-key');
const hk1 = hkdf(sha256, inputKey, salt, info, 32);

// == same as
import { extract, expand } from '@noble/hashes/hkdf.js';
const prk = extract(sha256, inputKey, salt);
const hk2 = expand(sha256, prk, info, 32);

Conforms to RFC 5869.

pbkdf2

import { pbkdf2, pbkdf2Async } from '@noble/hashes/pbkdf2.js';
import { sha256 } from '@noble/hashes/sha2.js';
const pbkey1 = pbkdf2(sha256, 'password', 'salt', { c: 524288, dkLen: 32 });
const pbkey2 = await pbkdf2Async(sha256, 'password', 'salt', { c: 524288, dkLen: 32 });
const pbkey3 = await pbkdf2Async(sha256, Uint8Array.from([1, 2, 3]), Uint8Array.from([4, 5, 6]), {
  c: 524288,
  dkLen: 32,
});

Conforms to RFC 8018.

scrypt

import { scrypt, scryptAsync } from '@noble/hashes/scrypt.js';
const scr1 = scrypt('password', 'salt', { N: 2 ** 16, r: 8, p: 1, dkLen: 32 });
const scr2 = await scryptAsync('password', 'salt', { N: 2 ** 16, r: 8, p: 1, dkLen: 32 });
const scr3 = await scryptAsync(Uint8Array.from([1, 2, 3]), Uint8Array.from([4, 5, 6]), {
  N: 2 ** 17,
  r: 8,
  p: 1,
  dkLen: 32,
  onProgress(percentage) {
    console.log('progress', percentage);
  },
  maxmem: 128 * 8 * (2 ** 17 + 1 + 1), // 128 * r * (N + p + 1)
});

Conforms to RFC 7914, Website

• N, r, p are work factors. It is common to only adjust N, while keeping r: 8, p: 1. See the blog post. JS doesn't support parallelization, making increasing p meaningless.
• dkLen is the length of output bytes e.g. 32 or 64
• onProgress can be used with async version of the function to report progress to a user.
• maxmem prevents DoS and is limited to 1GB + 1KB (2**30 + 2**10), but can be adjusted using formula: 128 * r * (N + p + 1)

Time it takes to derive Scrypt key under different values of N (2**N) on Apple M4 (mobile phones can be 1x-4x slower):

[!NOTE] We support N larger than 2**20 where available, however, not all JS engines support >= 2GB ArrayBuffer-s. When using such N, you'll need to manually adjust maxmem, using formula above. Other JS implementations don't support large N-s.

argon2

import { argon2d, argon2i, argon2id } from '@noble/hashes/argon2.js';
const arg1 = argon2id('password', 'saltsalt', { t: 2, m: 65536, p: 1, maxmem: 2 ** 32 - 1 });

Argon2 RFC 9106 implementation.

[!WARNING] Argon2 can't be fast in JS, because there is no fast Uint64Array. It is suggested to use Scrypt instead. Being 5x slower than native code means brute-forcing attackers have bigger advantage.

eskdf

import { eskdf } from '@noble/hashes/eskdf.js';
const kdf = await eskdf('example-university', 'beginning-new-example');
console.log(kdf.fingerprint);
const key = kdf.deriveChildKey('aes', 0);
kdf.expire();

Experimental KDF for deriving application-specific child keys from a username + password pair, built on scrypt, pbkdf2 and hkdf with fixed work factors. Non-standard: prefer scrypt or argon2 for new designs.

webcrypto: friendly wrapper

import { sha256, sha384, sha512, hmac, hkdf, pbkdf2 } from '@noble/hashes/webcrypto.js';
import { randomBytes } from '@noble/hashes/utils.js';
const whash = await sha256(Uint8Array.from([0xca, 0xfe, 0x01, 0x23]));

const key = new Uint8Array(32).fill(1);
const msg = new Uint8Array(32).fill(2);
const wmac = await hmac(sha256, key, msg);

const inputKey = randomBytes(32);
const salt = randomBytes(32);
const info = new TextEncoder().encode('application-key');
const hk1 = await hkdf(sha256, inputKey, salt, info, 32);

const pbkey1 = await pbkdf2(sha256, 'password', 'salt', { c: 524288, dkLen: 32 });

Sometimes people want to use built-in crypto.subtle instead of pure JS implementation. However, it has terrible API.

We simplify access to built-ins with API which mirrors noble-hashes. The overhead is minimal - just 30+ lines of code, which verify input correctness.

[!NOTE] Webcrypto methods are always async.

utils

import { bytesToHex as toHex, randomBytes } from '@noble/hashes/utils.js';
console.log(toHex(randomBytes(32)));

• bytesToHex will convert Uint8Array to a hex string
• randomBytes(bytes) will produce cryptographically secure random Uint8Array of length bytes

Security

The library has been audited:

• at version 2.2.0, in Apr 2026, by ourselves (self-audited)
  • Scope: everything
  • Changes since audit
• at version 1.0.0, in Jan 2022, independently, by Cure53
  • PDFs: website, in-repo
  • Scope: everything, besides blake3, sha3-addons, sha1 and argon2, which have not been audited
  • The audit has been funded by Ethereum Foundation with help of Nomic Labs

It is tested against official (ACVP / KAT) vectors, cross-library chained hashing, sliding-window length sweeps and property-based tests (fast-check), and is being fuzzed in the separate repo.

If you see anything unusual: investigate and report.

Constant-timeness

We're targetting algorithmic constant time. JIT-compiler and Garbage Collector make "constant time" extremely hard to achieve timing attack resistance in a scripting language. Which means any other JS library can't have constant-timeness. Even statically typed Rust, a language without GC, makes it harder to achieve constant-time for some cases. If your goal is absolute security, don't use any JS lib — including bindings to native ones. Use low-level libraries & languages.

Memory dumping

The library shares state buffers between hash function calls. The buffers are zeroed-out after each call. However, if an attacker can read application memory, you are doomed in any case:

• At some point, input will be a string and strings are immutable in JS: there is no way to overwrite them with zeros. For example: deriving key from scrypt(password, salt) where password and salt are strings
• Input from a file will stay in file buffers
• Input / output will be re-used multiple times in application which means it could stay in memory
• await anything() will always write all internal variables (including numbers) to memory. With async functions / Promises there are no guarantees when the code chunk would be executed. Which means attacker can have plenty of time to read data from memory
• There is no way to guarantee anything about zeroing sensitive data without complex tests-suite which will dump process memory and verify that there is no sensitive data left. For JS it means testing all browsers (incl. mobile), which is complex. And of course it will be useless without using the same test-suite in the actual application that consumes the library

Supply chain security

• Commits are signed with PGP keys to prevent forgery. Be sure to verify the commit signatures
• Releases are made transparently through token-less GitHub CI and Trusted Publishing. Be sure to verify the provenance logs for authenticity.
• Rare releasing is practiced to minimize the need for re-audits by end-users.
• Dependencies are minimized and strictly pinned to reduce supply-chain risk.
  • We use as few dependencies as possible.
  • Version ranges are locked, and changes are checked with npm-diff.
• Dev dependencies are excluded from end-user installs; they’re only used for development and build steps.

For this package, there are 0 dependencies; and a few dev dependencies:

• jsbt contains helpers for building, benchmarking & testing secure JS apps. It is developed by the same author
• prettier, fast-check and typescript are used for code quality / test generation / ts compilation

Randomness

We rely on the built-in crypto.getRandomValues, which is considered a cryptographically secure PRNG.

Browsers have had weaknesses in the past - and could again - but implementing a userspace CSPRNG is even worse, as there’s no reliable userspace source of high-quality entropy.

Quantum computers

Cryptographically relevant quantum computer, if built, will allow to utilize Grover's algorithm to break hashes in 2^n/2 operations, instead of 2^n.

This means SHA256 should be replaced with SHA512, SHA3-256 with SHA3-512, SHAKE128 with SHAKE256 etc.

Australian ASD prohibits SHA256 and similar hashes after 2030.

Upgrading

Supported node.js versions:

• v2: v20.19+ (ESM-only)
• v1: v14.21+ (ESM & CJS)

v2.0 changelog:

• The package is now ESM-only. ESM can finally be loaded from common.js on node v20.19+
• .js extension must be used for all modules
  • Old: @noble/hashes/sha3
  • New: @noble/hashes/sha3.js
  • This simplifies working in browsers natively without transpilers
• Only allow Uint8Array as hash inputs, prohibit string
  • Strict validation checks improve security
  • To replicate previous behavior, use utils.utf8ToBytes
• Rename / remove some modules for consistency. Previously, sha384 resided in sha512, which was weird
  • sha256, sha512 => sha2.js (consistent with sha3.js)
  • blake2b, blake2s => blake2.js (consistent with blake3.js, blake1.js)
  • ripemd160, sha1, md5 => legacy.js (all low-security hashes are there)
  • _assert => utils.js
  • crypto internal module got removed: use built-in WebCrypto instead
• Improve typescript types & option autocomplete
• Bump compilation target from es2020 to es2022

Contributing & testing

npm install && npm run build && npm test will build the code and run tests.

There are additional slow suites: timing-based DoS tests npm run test:dos, multi-hour large-input tests npm run test:slow, ACVP LDT vectors npm run test:acvp, and memory-intensive KDF tests npm run test:ultra. The 9–17GiB scrypt cases require an explicitly provisioned machine and run separately with npm run test:ultra:scrypt.

test/misc directory contains unrolled implementations (sha3, argon2) and misc helper scripts.

Some hashes are outside of scope of the library:

• Pedersen in micro-zk-proofs
• Poseidon in noble-curves
• Poly1305 & GHash in noble-ciphers

See paulmillr.com/noble for useful resources, articles, documentation and demos related to the library.

Speed

npm run benchmark

Benchmarks measured on Apple M4.

The library could be 3x faster by utilizing loop unrolling. It isn't used because unrolling a) would increase bundle size b) make lib un-readable c) current perf is "fast enough" for most use-cases.

If you need truly exemplar performance, switch to awasm-noble, which does unrolling in an auditable way and allows to achieve 10GB/s BLAKE3.

# 32B
sha256 x 2,178,649 ops/sec @ 459ns/op
sha512 x 783,699 ops/sec @ 1μs/op
sha3_256 x 520,562 ops/sec @ 1μs/op
sha3_512 x 526,870 ops/sec @ 1μs/op
kt128 x 762,195 ops/sec @ 1μs/op
kt256 x 769,230 ops/sec @ 1μs/op
turboshake128 x 902,527 ops/sec @ 1μs/op
blake256 x 796,812 ops/sec @ 1μs/op
blake2b x 455,788 ops/sec @ 2μs/op
blake2s x 1,015,228 ops/sec @ 985ns/op
blake3 x 1,095,290 ops/sec @ 913ns/op
ripemd160 x 1,594,896 ops/sec @ 627ns/op
md5 x 2,123,142 ops/sec @ 471ns/op
sha1 x 1,876,172 ops/sec @ 533ns/op
hmac(sha256) x 500,000 ops/sec @ 2μs/op
hmac(sha512) x 188,857 ops/sec @ 5μs/op
kmac256 x 149,633 ops/sec @ 6μs/op
blake3(key) x 957,854 ops/sec @ 1μs/op

# 1MB
sha256 x 293 mib/sec
sha512 x 128 mib/sec
sha3_256 x 76.7 mib/sec
sha3_512 x 40.2 mib/sec
kt128 x 177 mib/sec
kt256 x 147 mib/sec
turboshake128 x 187 mib/sec
blake256 x 57.3 mib/sec
blake2b x 66.9 mib/sec
blake2s x 79.6 mib/sec
blake3 x 100 mib/sec
ripemd160 x 179 mib/sec
md5 x 273 mib/sec
sha1 x 401 mib/sec
hmac(sha256) x 290 mib/sec
hmac(sha512) x 127 mib/sec
kmac256 x 74.8 mib/sec
blake3(key) x 97.2 mib/sec

## KDF
hkdf(sha256) x 242,541 ops/sec @ 4μs/op
blake3(context) x 495,294 ops/sec @ 2μs/op
pbkdf2(sha256, c: 2 ** 18) x 4 ops/sec @ 205ms/op ± 11.57% (203ms..206ms)
scrypt(n: 2 ** 19, r: 8, p: 1) x 1 ops/sec @ 768ms/op
argon2id(t: 1, m: 128MB) x 2 ops/sec @ 357ms/op ± 26.23% (349ms..364ms)

License

The MIT License (MIT)

Copyright (c) 2022 Paul Miller (https://paulmillr.com)

See LICENSE file.