How Dropbox securely stores your passwords

It’s universally acknowledged that it’s a bad idea to store plain-text passwords. If a database containing plain-text passwords is compromised, user accounts are in immediate danger. For this reason, as early as 1976, the industry standardized on storing passwords using secure, one-way hashing mechanisms (starting with Unix Crypt). Unfortunately, while this prevents the direct reading of passwords in case of a compromise, all hashing mechanisms necessarily allow attackers to brute force the hash offline, by going through lists of possible passwords, hashing them, and comparing the result. In this context, secure hashing functions like SHA have a critical flaw for password hashing: they are designed to be fast. A modern commodity CPU can generate millions of SHA256 hashes per second. Specialized GPU clusters allow for calculating hashes at a rate of billions per second.

Over the years, we’ve quietly upgraded our password hashing approach multiple times in an ongoing effort to stay ahead of the bad guys. In this post, we want to share more details of our current password storage mechanism and our reasoning behind it. Our password storage scheme relies on three different layers of cryptographic protections, as the figure below illustrates. For ease of elucidation, in the figure and below we omit any mention of binary encoding (base64).

layersMultiple layers of protection for passwords

We rely on bcrypt as our core hashing algorithm with a per-user salt and an encryption key (or global pepper), stored separately. Our approach differs from basic bcrypt in a few significant ways.

First, the plaintext password is transformed into a hash value using SHA512. This addresses two particular issues with bcrypt. Some implementations of bcrypt truncate the input to 72 bytes, which reduces the entropy of the passwords. Other implementations don’t truncate the input and are therefore vulnerable to DoS attacks because they allow the input of arbitrarily long passwords. By applying SHA, we can quickly convert really long passwords into a fixed length 512 bit value, solving both problems.

Next, this SHA512 hash is hashed again using bcrypt with a cost of 10, and a unique, per-user salt. Unlike cryptographic hash functions like SHA, bcrypt is designed to be slow and hard to speed up via custom hardware and GPUs. A work factor of 10 translates into roughly 100ms for all these steps on our servers.

Finally, the resulting bcrypt hash is encrypted with AES256 using a secret key (common to all hashes) that we refer to as a pepper. The pepper is a defense in depth measure. The pepper value is stored separately in a manner that makes it difficult to discover by an attacker (i.e. not in a database table). As a result, if only the password storage is compromised, the password hashes are encrypted and of no use to an attacker.

Why not use {scrypt, argon2} over bcrypt?

We considered using scrypt, but we had more experience using bcrypt. The debate over which algorithm is better is still open, and most security experts agree that scrypt and bcrypt provide similar protections.

We’re considering argon2 for our next upgrade: when we moved to our current scheme, argon2 hadn’t (yet) won the Password Hashing Competition. Additionally, while we believe argon2 is a fantastic password hashing function, we like that bcrypt has been around since 1999 without any significant vulnerabilities found.

Why is the global pepper used for encryption instead of hashing?

Recall that the global pepper is a defense in depth measure and we store it separately. But storing it separately also means that we have to include the possibility of the pepper (and not the password hashes) being compromised. If we use the global pepper for hashing, we can’t easily rotate it. Instead, using it for encryption gives us similar security but with the added ability to rotate. The input to this encryption function is randomized, but we also include a random initialization vector (IV).

Going forward, we’re considering storing the global pepper in a hardware security module (HSM). At our scale, this is an undertaking with considerable complexity, but would significantly reduce the chances of a pepper compromise. We also plan to increase our bcrypt strength in our next update.

Moving forward

We believe this use of SHA512, plus bcrypt, and AES256 is currently among the strongest and most future-proof methods to protect passwords. At the same time, we know that attackers are continuously evolving—and our defenses will too. Our password hashing procedure is just one of many measures we use to secure Dropbox. We’ve deployed additional safeguards against online brute-force attacks like rate-limiting password attempts, captchas, and a range of abuse mitigations. Like the diagram above, there are many layers to maintaining robust security, and we’re actively investing in all of them. We’d love to hear your thoughts.