NLS: Warcraft III's password system
![]()
The New Logon System proves a player knows their password without ever sending it or anything that could replace it. It's a modified SRP-6.
Warcraft III logs on with NLS (logon type 0x02 in SID_AUTH_INFO). The server stores a salt and a verifier for each account, never the password or a hash that could stand in for it. The messages are SID_AUTH_ACCOUNTCREATE, SID_AUTH_ACCOUNTLOGON and SID_AUTH_ACCOUNTLOGONPROOF.
Constants
| Name | Value |
|---|---|
g |
47 (0x2F) |
N |
0xF8FF1A8B619918032186B68CA092B5557E976C78C73212D91216F6658523C787 (256 bits) |
| Hash | Standard SHA-1 everywhere. Not X-SHA-1. |
The calculation
C is the account name, P the password, and ‖ means "followed by".
x = SHA1( s ‖ SHA1( UPPER(C) ‖ ":" ‖ UPPER(P) ) ) read as a little-endian number
v = g^x mod N the verifier, stored with s
A = g^a mod N client; a is random for each logon
B = (v + g^b mod N) mod N server; b is random for each logon
u = first 4 bytes of SHA1(B) read as a BIG-endian number; never sent
S = client: (B − v)^(a + u·x) mod N
server: (A · v^u)^b mod N both sides get the same S
K = SHA1 of S's even-numbered bytes and SHA1 of its odd-numbered bytes,
interleaved byte by byte 40 bytes
I = SHA1(g) XOR SHA1(N) a 20-byte constant
M1 = SHA1( I ‖ SHA1(UPPER(C)) ‖ s ‖ A ‖ B ‖ K ) the client's proof
M2 = SHA1( A ‖ M1 ‖ K ) the server's proof
Byte order: where implementations go wrong
- Every number is 32 bytes, little-endian, padded with zeros:
s,v,AandBon the wire, and whenever they are hashed.Kis always 40 bytes. - Don't drop leading zero bytes. Encoding a number at its natural length passes your own tests, but about one logon in 256 has a zero top byte, hashes a byte short, and fails. The result is "some accounts never work".
- Uppercase the name and password, in
xand inM1. (Classic games lowercase instead; NLS is the opposite.) xis read little-endian, butuis read big-endian. Both are correct.IhashesNin its little-endian form. Its bytes are6C0E97ED 0AF96BAB B15889EB 8BBA25A4 F08C01F8.- NLS version 1 (logon type
0x01, older clients) reverses the byte order ofN.
A useful shortcut: if the server accepts the client's M1, its M2 is automatically right. Both are built from the same A and K, so "the client rejects M2" is never a math bug. If the client disconnects after a successful proof, check the reply's length instead (see SID_AUTH_ACCOUNTLOGONPROOF).
Test values
For account Tagban, password hunter2, salt bytes 00 01 02 … 1F, a = bytes 01 02 … 20 and b = bytes 02 03 … 21 (both little-endian):
x = 0x395e620c90a3919c75ede2fdd52aaf8cd6b28d39
v = 7f2bf7dcd443e0f945a5a7bcbfec2f5e3fee8242d09b62aac46c194963dcafef
A = f085e0261e42a247db6e0bb57ae40d56746b0c16783606718a2c500d3fed2594
B = 423ac3a9427c556639f31804d59e3e2c409e182add1648440ed88622abb94580
u = 0x672cb5c0
K = 2eeb577075dfcbdd4ec5c17e13c8df4efc8bc573ce35f665c6a913bf0c6e2febf0b0c23a00428941
M1 = a4d99ac492f45207ce0ed8e1c9729f38a425d151
M2 = ba079a3bfcf36d706f1df1dd2934c3252aad086f
These come from a reference model whose client and server sides agree over random inputs. A real Warcraft III client has since logged on successfully with the same implementation.
Server safety checks
The game doesn't require these, but a server should still refuse A mod N = 0, pick a fresh b for every logon, compare M1 in constant time, and never reuse b after a failure.
Sources and credits
- Command Center: WARCRAFT3.md: the algorithm, byte-order rules and test values, checked with an independent Python model
- Command Center: WARCRAFT3-FIELD-NOTES.md: the two mistakes that pass your own tests but fail a real client
- BNETDocs, NLS/SRP Protocol: written by Sixen, and edited 11 times since.
This page is written in our own words. BNETDocs is credited for its research; none of its text is copied. See Credits.