The Pwned buzz and why you really don't need this database


If you watched your social media channels over the last few days, you probably haven’t missed the media buzz about the launch of the new Pwned V2 Passwords database, not least because of our popular competitor 1Password integrating it as a service for you, so you can check if your beloved passwords have been compromised.

The database contains the SHA-1 hashes of half a billion leaked passwords. So if you want a new and secure password, you can now easily check it against a fairly large collection of known bad and compromised passwords. Neat, right?

Well…, we do appreciate the ’;–have i been pwned? service to check if one of your accounts has been stolen immensely, but let me explain why you really don’t need a database of bad passwords.

A few quick tests

For testing purposes, I downloaded the database dump, which has a compressed size of about 8GB and decompresses to 30GB of plain text. Each line in the file contains the SHA-1 hash of a password and usage counts.

A quick first test of the top 10 most-popular passwords of 2017 gave no surprising results. All 10 were found fairly easily towards the top of the file. But here’s the fun fact: adding whitespace to the end of the passwords more often than not fails to return results. " 12345678" can be found in line 4778908, but “12345678 " does not exist.

So here we directly have our problem: despite the (seemingly) huge mass of half a billion passwords, it’s surprisingly trivial to create password variants that are not in the database. Neither of “password “, “pass>word”, or “pass}word” could be found and they are all within edit distance of 1 to the original “password”. Permutations are also reasonably effective. “dRowssap” is “password” spelled backwards with a capitalized R and it cannot be found. People are creative when they have to update their passwords and many variants are in the database, but you will always find a very simple one that isn’t.

The math

The math behind this is rather simple. Assume, we have a password with only lowercase and uppercase letters and numbers. This gives us an alphabet size of 62 (26 + 26 + 10). With an 8-character password, we end up with 62^8 = 218,340,105,584,896 possible permutations (with repetition) of these 8 characters, which is already a lot more than the 501,636,842 passwords contained in the database. So with any random 8-character combination of letters and numbers we have a chance of roughly ~2.3 × 10^-6 of finding it in the database.

Password managers to the rescue

With chances of finding trivial 8-character passwords in the database already being so low, here comes the paradox part: How does a password manager, whose one and only purpose is to securely store arbitrarily many and complex passwords, so you as a human being don’t have to remember them, find it useful to check your password against half a billion of those passwords?

There are only two answers for this:

  1. You don’t use your password manager to generate secure passwords (start doing that now!)
  2. You are using services that got compromised and didn’t hash their passwords properly (stop using them!)

The second point probably needs a little explanation. When you create an account at an online service, your password (hopefully) isn’t stored in plain text. Instead, a hash function is run on it (hopefully one that is slow) to create a fixed-length representation of your password. This representation is always the same for the same input password. To make it impossible to pre-calculate a large number of hashes and compare them, a so-called salt is added to the password before hashing, which is stored in plaintext next to the hash. So next time, you log in, the service takes your password, appends the salt, runs it through the hash function and checks if the output matches the one stored in the user database.

So what happens if the user database is leaked? Well, don’t worry too much, the attacker can only get the salted hash and not your actual password. Unless…, the service didn’t do their homework and stored your password in plain text nonetheless, in which case it will eventually end up in one of those Pwned databases.

Complex passwords and brute force

Now let’s assume an attacker got hold of your password’s hash. They can still try to guess the correct password by brute force. The amount of time needed depends largely on two factors:

  • the number of possible password candidates
  • speed and memory requirements of the password hashing scheme

The number of possible password candidates is controlled by the complexity of your password. The other factor is probably out of your reach when you are using an online service. It determines how many guesses an attacker can make per second. You want this to be slow. This is why we use Argon2 in KeePassXC 2.3 (or the older AES-KDF for KDBX 3.1 in KeePassXC 2.2).

A plain SHA-1 hash is fast and therefore not a good password hashing algorithm (and it also hasn’t aged well for other reasons).

Benchmarks suggest that an expensive and beefy modern GPU cluster can perform about 8.5 billion SHA-1 hashes per second (!). Such a cluster could generate all hashes of the whole Pwned database in 0.6 seconds.

If we take our original 8-character password, it could generate all possible passwords in a little more than seven hours. And this is the worst-case estimate. On average, an attacker will find the correct password after 50% of the time, which is 3.5 hours. So take this into account the next time you think about sending your SHA-1 password hash to some online password database (even if it’s just the first 5 characters of the hash).

So what if we just increase our password size? We use a damn password manager, so why not just dial it up to 30 characters? This will create 62^30 ≈ 5.9 × 10^53 possible candidates. Testing all these at 8.5 billion hashes per second would take a rough 2.2 × 10^36 years. That’s about 159,594,201,897,441,234,244,091,519 times the age of the universe (take half of that to get from worst case to average, I’m being generous). I guess you can see that the chances of one of those passwords being in the Pwned database are very, veeeery low and that’s only with letters and numbers. If we take the full ASCII set (excluding the NULL character), we have 127^30 possible candidates, which is not quite the number of atoms in the universe, but at least Archimedes’ estimate of how many grains of sand would fill it.

Summary

If you made it to this point, then congratulations. Let me sum things up.

Checking your (strong) password against a database is only useful if (and that’s a very big if) your account was hacked and the service didn’t hash your password properly. But in that case, it’s almost certain that your username was leaked, too, so better check that against a database. It’s simply the more logical thing to do. If your account was hacked, but the service hashed and salted your password, you wouldn’t find it in a password database anyway, even if it were in there. On the other hand, you will always be able to find your username.

Feedback

Please report any bugs you encounter at our GitHub issue tracker. We are also available on Matrix for real-time feedback and discussions. See our contact page for further options.