The current NIST password rules are in SP 800-63B Revision 4, finalised on 31 July 2025 and published that August. It replaces the March 2020 edition and it is not a set of suggestions: it is the federal standard, and most enterprise policies still run on the rules it just deleted.
Almost every page ranking for this query pastes the same summary table. The part they leave out is that SP 800-63B is written for the verifier — the site that stores your password — not for you, the person choosing one. Those are two different checklists, so this page gives both, plus the number behind the 15.
What SP 800-63B-4 requires
All of it lives in section 3.1.1.2 ("memorized secrets"). SHALL is a requirement, SHOULD is a recommendation.
| Topic | What Revision 4 says | What it means in practice |
|---|---|---|
| Minimum length, password alone | SHALL be at least 15 characters | If a password by itself gets someone in, the floor is 15 |
| Minimum length, inside MFA | SHALL be at least 8 characters | Shorter is allowed only when every account also carries a second factor |
| Maximum length | SHOULD accept at least 64 characters | Do not cap a manager-generated string or a passphrase |
| Character set | SHOULD accept printing ASCII, the space, and Unicode | Spaces make passphrases possible; rejecting accents locks people out |
| Composition rules | SHALL NOT impose them | No "one capital, one digit, one symbol" requirements |
| Expiry | SHALL NOT require periodic change | No 90-day rotation. Change only on evidence of compromise |
| Blocklist | SHALL check new and changed passwords against commonly used, expected and compromised values | The rule that actually prevents Summer2026! |
| Hints and security questions | SHALL NOT store hints or prompt for knowledge-based answers | "First pet's name" is out as a recovery path |
| Password managers | SHALL allow them and autofill; SHOULD allow paste | Blocking paste blocks the best password storage most people have |
| Truncation | SHALL verify the entire password | A silent 20-character cut turns a strong password into a collision |
| Failed attempts | SHALL rate-limit | Online guessing is a rate problem, not a length problem |
| Storage | Salt of at least 32 bits, salted hash stored as a form resistant to offline attack | Length does nothing if the database hands out the answer |
What actually changed from the 2020 revision
- The floor moved from 8 to 15 when a password is the only authenticator. Eight survives only inside MFA.
- Composition rules went from discouraged to prohibited. Demanding a symbol now means rejecting a configuration that NIST says is worse than a longer plain one.
- Periodic rotation is prohibited, not merely discouraged, with the exception of evidence of compromise.
- Breach screening is a requirement at creation and at change, not an optional add-on.
- Passkeys are named in the framework: synced passkeys count at AAL2 and device-bound at AAL3, which puts the WebAuthn path on the same footing as a hardware key.
- SMS OTP stays restricted. Nothing in Revision 4 rehabilitates a code sent over a mobile network.
If you run the login form
This is the part that gets traded away as "we'll fix it next sprint". Each line below is a requirement from the table above, expressed as a change you can make in an afternoon.
# 1. Length, not decoration
<input type="password" name="password"
minlength="15" maxlength="64"
autocomplete="new-password" autofocus>
# 2. Do not block paste; do not strip spaces or Unicode
# 3. Never truncate before hashing - check the whole value
# 4. Store with a salted, slow KDF (Argon2id, scrypt, bcrypt or PBKDF2)
# 5. Rate-limit failures per account AND per IP
# 6. Screen against a breach corpus at signup and at change
For screening, the k-anonymity pattern matters: you hash the candidate, send only the first five characters of the hash to a breach API, and compare the returned suffix list locally. The password never leaves your server, and the API never learns it. See k-anonymity for why that is not a downgrade in security.
Two habits are worth deliberately breaking. Do not send a reminder email with the password in it — NIST forbids hints and that is a hint with better delivery. And do not expire a password on a calendar; expire it when your breach feed says the account was in a dump, which is the same signal but arrives when it matters.
If you are just choosing a password
You cannot make a site follow the standard, but you can make every password you create clear it, and that is easier than the old rules implied.
- Let the manager generate it. 15 to 20 random characters, unique per site. You never type it, so length costs you nothing.
- Use a passphrase for the few you must type. Four to six unrelated words with spaces. "correct horse battery staple" is long, allowed by NIST, and easy to remember; the passphrase generator draws from a proper word list.
- Two factors on anything that holds money, mail or identity, preferring an authenticator app or a security key over SMS. If SMS is all a service offers, it still beats nothing.
- Stop rotating. A password that has not leaked is not made safer by being changed; it is made more likely to be written on a sticky note.
- Check exposure, not vibes — put the address behind your login into a breach check rather than assuming that a strong password was never leaked from the site.
Why 15, and what that number buys
Fifteen is not a magic value; it is the point where a password in a realistic character set stops being cheap to attack at the speed of a GPU against a fast hash.
| Password shape | Search space | Time at 1011 guesses/second |
|---|---|---|
| 8 characters, full printable ASCII | about 52 bits | minutes to hours |
8 characters with the usual substitutions (P@ssw0rd) | far less than 52 bits, because rules models know the pattern | seconds |
| 15 random lowercase letters | about 70 bits | centuries |
| 20 random characters from a manager | about 130 bits | beyond brute force at any scale |
| 5 random words from a 7,776-word list | about 65 bits | centuries, and you can remember it |
The honest caveat: those figures assume a fast hash such as unsalted MD5 or NTLM. Against bcrypt, scrypt or Argon2id every guess costs real time, so an 8-character password behind a good KDF is not equal to the 19 hours in the table. Length stacks on top of the KDF rather than replacing it, which is exactly why NIST requires both.
The reason composition rules were removed is visible in the second row: humans satisfy a symbol requirement in predictable ways, and the attack models already contain those ways. Twenty years of P@ssw0rd1 is the evidence. What a blocklist does is refuse the predictable value outright, and what length does is make the unpredictable value expensive. Neither needs a capital letter.
Common misreads
- "NIST says 8 characters" — only when a second factor always applies. A verified second factor is not the same as a checkbox labelled "remember this device".
- "NIST banned complexity" — it banned requiring character classes. A long password with symbols is fine; it just no longer has to look a certain way.
- "NIST says passwords must expire after 60 days" — the opposite. Rotation is prohibited except on compromise.
- "This is US-only" — it is US federal, but auditors elsewhere cite it, and most of the technical content matches NCSC and CISA advice.
- "64 characters is the maximum allowed" — 64 is the minimum maximum a verifier should support. Accepting more is not a violation.
Frequently asked
Is there a 2026 revision of SP 800-63B?
No. The 2026 guidance is still Revision 4, published in 2025, and there is no successor in final form. Any page claiming a new 2026 document is usually re-dating the 2025 one.
What is the exact NIST minimum password length?
15 characters when the password is the only authenticator; 8 when it is one factor inside multi-factor authentication. Verifiers should also accept at least 64.
Do I still need to change my password every 90 days?
Not for policy reasons. Change it when there is evidence of compromise — a breach notice, a suspicious sign-in, a reused password appearing in a dump — and not because a quarter ended.
Are special characters pointless now?
No, they add entropy. What is pointless is demanding them, because the demand is what produces predictable patterns. Choose length first; add symbols if you can remember them.
Does NIST approve password managers?
It requires verifiers to allow them and to accept autofill, and recommends allowing paste. Blocking a manager is now a compliance problem, not a hard-nosed security choice.
Where does MFA fit?
It is the only thing that makes the 8-character allowance legitimate, and it is the difference between a password that leaks and an account that is taken over. If you are picking one upgrade this year, pick a second factor over a longer password.