pboyd.io

How Not to Store Passwords

Here’s a fun list to look through: Dumb Password Rules. Most of the rules seem arbitrary, like only allowing digits, but some hint at deeper problems. For instance, preventing single-quotes. They aren’t inserting passwords into a database without a SQL placeholder, right? Nearly every site on that list has a needlessly short maximum password size. If they’re storing passwords correctly, there’s no need for this. This post will go through a few bad ways to store a password and you can see what I mean.

Plain text

The absolute simplest way to store a password in a database is to just store it in plain text. Here’s an example, it’s probably nothing surprising. Here’s the code to create an account:

def create_account(self, cursor, username, password):
    if len(password) < 8:
        raise PasswordTooShort()

    if len(password) > 16:
        raise PasswordTooLong()

    cursor.execute("SELECT count(*) FROM plaintext WHERE username=%s", (username,))
    result = cursor.fetchone()
    if result[0] > 0:
        raise UsernameTaken()

    cursor.execute("INSERT INTO plaintext (username, password) VALUES (%s, %s)", (username, password))

Nothing complicated here, just a few basic checks and an INSERT. My database column has a max size of 16 characters, which I had enforce to prevent the user getting an ugly 500 error1.

Of course, this is terrible. Even if the site in question doesn’t have anything of real value, the passwords themselves can be valuable. A lot of people use the same password for everything. When a sites asks for a password, they put in their password. The password they gave you is the same password they’ve used for their email and bank account. Don’t feel too special, they probably give it out to anyone that asks for the WiFi password.

I thought everyone knew not to store passwords in plain text, but I was wrong. Have I Been Pwned is another fun list to look through. There are plenty of sites that leaked plain text passwords. From big names that you’d think would know better2.

Encryption

So you’ve decided not to store your passwords in plain text. Good for you. If you think about it, a password is kind of like a secret message, so why not encrypt it? The code is similar to the plain text version, except it calls a library to encrypt and decrypt the password.

key = b'this is a super-duper secret key'

def encrypt_password(self, plaintext):
    box = nacl.secret.SecretBox(self.key)
    return box.encrypt(bytes(plaintext, 'utf-8')).hex()

def decrypt_password(self, ciphertext):
    box = nacl.secret.SecretBox(self.key)
    bc = bytes.fromhex(ciphertext)
    return box.decrypt(bc).decode('utf-8')

This is an improvement over plain text, and it would probably be difficult for an attacker without the key to decrypt a password. But, depending on how the encrypted passwords were obtained, the attacker may have the key.

Anyway, the mere existence of the key complicates things. Now we have to figure out how to store it, and who can have it, how the software that needs the key will get it, and a bunch of other problems besides. My little demo side-stepped all of this and just hard-coded the key. Somehow, though, the key has to be available in the login code, and that makes it possible for an attacker to get it.

Another problem is that the length of the password determines the length of the encrypted text. Which means we still have to enforce a max password size to avoid overflowing the database column. But also the encrypted text gives away the length of the password, or a least an approximate length. This makes it much easier to brute-force leaked passwords.

I worked on a (now dead) project that encrypted passwords like this, and I thought it was a unique “innovation”. In fact, I thought it was so rare that I wasn’t going to include it here. But I was wrong, Adobe did the exact same thing.

Hashes

Cryptographic hash functions allow one person to prove they know a secret without having to say what the secret is. These functions work by deriving a value from the password. There is no algorithm to get the input text back from the resulting hash. Which is exactly what you want for a password, since you don’t need to know the password, you just need to know if it’s correct. In fact, it’s a liability if you do know the password, so hash functions make sense.

There are a few hashing algorithms to choose from, and most programming languages provide implementations of the popular ones. My demo app has implementations for MD5 and SHA256. MD5 is largely broken, so it shouldn’t be used any more. The SHA-2 family is still considered secure.

Here’s the code to check a password:

def login(self, cursor, username, password):
    cursor.execute("SELECT password FROM sha256 WHERE username=%s", (username,))
    result = cursor.fetchone()
    if result == None:
        raise BadLogin()

    if result[0] != self.hash_password(password):
        raise BadLogin()

def hash_password(self, password):
    return hashlib.sha256(bytes(password, 'utf-8')).hexdigest()

It retrieves the hash from the database and compares it to a hash of the password from user. If the hashes match, the password is correct.

Hash functions always produce the same size output, so (within reason) there’s no need to limit the size of a password.

Unfortunately, a hash by itself is not enough. It’s pretty common that several accounts will have the same bad passwords (you didn’t think you were the only one who used superman, did you?) These stick out in password dumps:

 username |                             password                             
----------+------------------------------------------------------------------
 bob      | 5e884898da28047151d0e56f8dc6292773603d0d6aabbdd62a11ef721d1542d8
 sally    | 5e884898da28047151d0e56f8dc6292773603d0d6aabbdd62a11ef721d1542d8
 jim      | c43d1b57a27277d2846f6058d32cd3882885656612b48907f1e0fb93c8370e2b

Bob and Sally both used password. If an attacker cracks one password, every account with that password has been cracked. There are lists of common passwords on the internet. All an attacker has to do is hash the passwords in the list and find the accounts that match in the dump. It won’t crack every password, but it will probably crack enough, and in very little time.

Cryptographic hashes are fast to compute on normal hardware. A GPU can make the work go faster. Some of the algorithms are used in digital currency mining, which has prompted the development of custom hardware (ASICs) to compute them, and that’s faster still. So a brute-force attack that tries every combination is actually becoming reasonable, even against strong passwords. If computing the hashes it too much work, attackers can also download a rainbow table, which is a pre-computed list of hashed password.

It’s been known for a long time that plain hashes for passwords is pretty dumb, but there are still frequent dumps of passwords using nothing more than a basic hash. LinkedIn, for example, lost SHA1 hashes. The infamous Ashley Madison breech contained MD5 password hashes (in addition to bcrypt).

Salted Hashes

Since an attacker can download a rainbow table of standard hashes, what if you just added a little extra text to every password? So store hashes of MyCoolSite{password}, then attackers would need a unique rainbow table just for MyCoolSite. Better still, give every password it’s own bit of text, then an attacker needs a new rainbow table for every password. And duplicate passwords still produce unique hashes. That bit of text is a “salt”.

Here’s some code to generate a salted hash:

def hash_password(self, password, salt=None):
    if salt == None:
        salt = secrets.token_bytes(16)

    salted_password = salt + bytes(password, 'utf-8')
    return salt.hex() + hashlib.sha256(salted_password).hexdigest()

A random salt is chosen and prepended to the password before it’s hashed. The salt is needed to verify the password, and it isn’t a secret, so it needs to be stored. It doesn’t matter how the salt is stored, I prepended it to the hash, but any way is fine as long as it can be retrieved.

To check a salted password, recalculate the hash with the same salt and see if the hashes match:

def check_password(self, hashed, password):
    salt = bytes.fromhex(hashed[0:32])
    new_hashed = self.hash_password(password, salt)
    return new_hashed == hashed

Salted hashes are a big improvement. But the speed that hashes can be calculated makes brute-force attacks against even salted passwords quite reasonable.

Key Derivation Functions

Salted hashes would be fine if they just took longer to brute-force. And that’s the principle behind key derivation functions (KDFs). They require more compute time (and some require memory also), which requires an attacker to spend real money to crack them.

There are a few algorithms to choose from. PBKDF2 and bcrypt may be brute-forced before too long. scrypt and argon2 look like they’ll be with us longer. There are implementations for every major programming language.

Password storage is usually not a selling point, so I understand that it’s a tempting corner to cut, but it’s a risk. And it seems like an unnecessary one, when the difference between dumb password storage and good password storage is just a matter of calling the right library.

All the code used in this post is on github.


  1. I’m using Postgres, which raises an error. Other databases, such as old versions of MySQL, silently truncate the value. Wild conjecture of course, but that sounds suspiciously like what happens here. [return]
  2. Also Sony in 2012. Seems like these two groups should have been talking to each other. [return]