Are Password Strength Meters Accurate? Mostly No, and Here Is the Test That Proves It
Are password strength meters accurate? Take two passwords and run them through the same meter. "P@ssw0rd123!" lights up green. "carpet static pretzel invoke" gets a yellow shrug. The first one falls to a dictionary attack in seconds. The second one would take centuries of guessing. The meter got it exactly backwards, and that inversion is not a bug in one bad meter. It is how most meters work.
I first noticed this years ago on a signup form that refused my generated passphrase and demanded I add a capital letter and a symbol. The form was not measuring strength. It was measuring compliance with a checklist, and checklists are what attackers count on. When every user is forced to add a symbol, nearly everyone adds "!" at the end. When every user needs a digit, they add a year. A rule that everyone follows the same way stops being a defense and starts being a hint.
What most meters actually measure
A typical meter counts four things: length, uppercase letters, digits, and symbols. Each box ticked adds points. Nothing in that calculation asks the one question that matters: how was this password chosen? That question decides everything, because entropy comes from the process, not the characters on screen.
Here is the math that makes the point concrete. A four word passphrase drawn at random from the EFF diceware list of 7,776 words carries about 52 bits of entropy: each word contributes log2(7776), roughly 12.9 bits, times four. "P@ssw0rd123!" looks like 13 characters from a 94 character alphabet, which naive math calls 85 bits. But it is not drawn from that alphabet. It is a dictionary word plus the single most common substitution pattern plus the most common suffix pattern, and cracking tools try exactly that combination within their first few thousand guesses. Its real strength is closer to 20 bits. The meter reports 85. The attacker needs 20. That gap is the whole story.
NIST saw this coming. Special Publication 800-63B deprecated composition rules in 2017, the rules that require a mix of character types, because they push users into predictable patterns while adding memorization pain. A meter built on those same rules inherits the same flaw. It rewards the appearance of randomness while punishing actual randomness that happens to be all lowercase words.
So are password strength meters accurate?
It depends on which meter, and the split matters. The old checklist meters are not accurate in any sense worth trusting. They cannot distinguish a random string from a decorated dictionary word, and they will happily certify the decorated word as strong. If the meter on a site you use only checks character classes, treat its green bar as decoration.
The exception is meters built on guess estimation, above all the open source zxcvbn library from Dropbox. Instead of counting character types, zxcvbn looks for words, names, dates, keyboard walks, repeats, and common substitutions, then estimates how many guesses an attacker using those patterns would need. Feed it "P@ssw0rd123!" and it sees a dictionary word, a leet substitution, and a sequence, and scores it accordingly. Feed it four random diceware words and it finds no pattern to exploit, so the score reflects the real search space. That is a fundamentally different kind of tool, and it is the only kind I would trust.
Even zxcvbn has a blind spot worth knowing: it cannot see breach data. A password can be random, long, and pattern free, and still be a bad choice if it has appeared in a breach, because attackers try leaked passwords first. That is why the honest check has three parts, not one: generate the password randomly, run it through a guess estimation meter like zxcvbn, and check it against a breach database. Any one of the three alone leaves a hole.
My opinion, stated plainly: the green bar on most signup forms is worse than useless, because it teaches people that complexity equals safety and then punishes the safest format there is. A passphrase of random words looks boring to a checklist meter. Boring is the point. Attackers cannot guess boring at scale, because boring has no pattern to exploit.
If you want a passphrase that passes the honest test instead of the checklist, generate one here. Random words, drawn offline, nothing sent anywhere. Then run it through a zxcvbn based checker if you want a second opinion. Just do not let a green bar talk you into swapping a random word for "P@ssw0rd" with decorations.
Frequently asked questions
Why does my passphrase score lower than a shorter complex password on some meters?
Because the meter counts character types, not randomness. An all lowercase passphrase ticks one box no matter how many words it has, while a short password with upper, lower, digit, and symbol ticks four. The meter is scoring variety, and variety is not strength.
What is zxcvbn and why is it better?
zxcvbn is an open source password strength estimator that models how an attacker actually guesses: dictionary words, names, dates, keyboard patterns, repeats, and common substitutions. It estimates the number of guesses needed instead of counting character classes, so it catches decorated dictionary words that checklist meters pass.
Can a strength meter tell whether my password was randomly generated?
No meter can see the generation process, which is the real source of entropy. A meter can only inspect the finished string. That is why the honest workflow is to generate randomly first and use the meter only as a sanity check, never as the thing that decides whether a password is good.
Does NIST still recommend complexity rules?
No. NIST Special Publication 800-63B deprecated composition rules in 2017. Current guidance favors length, screening new passwords against breach lists, and allowing spaces and Unicode characters, which is exactly the direction passphrases point.
Should I trust a meter that says my password takes centuries to crack?
Only if it explains its model. A crack time estimate is only as good as the guess rate and attack model behind it. An offline attack against a leaked hash runs billions of guesses per second, while a rate limited login form allows a handful. A meter that shows one number without saying which attack it models is telling you half a story.
Related reading: How Many Words Does Your Passphrase Need? Entropy Math, Worked Out · Diceware Passphrase vs Random Password: Which Wins? · Is It Safe to Use an Online Password Generator? · How Long Does It Take to Crack a Passphrase?
One practical security guide a week. Get it by email.