Every time I read something from OWASP, it has a bunch of glaring deficiencies, and appears to be an uneditable wiki. This time around:
Authentication_Cheat_Sheet: password rules which only apply to US-ASCII. That means people with non-US keyboards may not even be able to set a password which meet the rules (eg. Θ and ẞ are upper case, but does not count as upper case in their suggested rules).
Password_Storage_Cheat_Sheet: misses the critical step of unicode canonicalisation and encoding before feeding the password into the slow one-way function (which invariably take an octet string, not textual strings). If you fail to do this, your system will spuriously reject correct passwords if the user logs in via different devices or input methods.
Password_Storage_Cheat_Sheet: '32-bit or 64b' length salt is certainly too short to be called cryptographically strong. Particularly, a 32-bit salt is not enough to avoid leaking password equality with good probability between users once you get past the birthday bound.
Password_Storage_Cheat_Sheet: implies that the caller is responsible for prepending the salt to the credential before inputting it into the slow one-way function. That's not how PBKDF2, scrypt or bcrypt work.
Your assertion about the salt is wrong. That's not what a salt is designed to prevent. A salt is to prevent pregenerated tables from being meaningful, therefore 32-64bit is plenty.
'good probability' is a little dramatic too. The birthday paradox would only apply if every user had the same password to begin with. With 32 bits, you would need about 77000 users with the exact same password to get to 50% probability of two having the same salt/hash. 2000 users with the same password is 0.1% probability of two being the same.
If you are concerned about equalities because you have thousands of users with the same password, input the username into the hash generation function as well. Don't waste space with needlessly large salts.
Let's consider Adobe, since we have actual data. Of 130,324,429 users, we know 1,911,938 chose the password '123456'. Consider just that population of users.
Assume a 64-bit salt was chosen uniformly at random for each user. The probability that this cryptosystem is catastrophically broken (i.e. leaks password equality between users in just this population) is about 2^-23. That's not good -- usually cryptosystems are designed to have a probability of failure between 2^-80 and 2^-256. Now multiply that up for each population of users choosing the same password, and you're in trouble.
Adding in other user-unique data is definitely a good idea!
> is about 2^-23. That's not good -- usually cryptosystems are designed to have a probability of failure between 2^-80 and 2^-256.
The failure cases are different though. Cryptosystems that require a strength of 80 to 256 bits are measuring their strength against an adversary brute-forcing or performing some other active discovery.
In this case, it's just the probability that two users will pick the same password and get the same salt. The attacker doesn't get to run many guesses, the database is just leaked 'as-is' and the hashes either show equality or they don't. So there is only a 1 in ~8mil chance that two in that group will show equality. Not bad for such a large population.
It doesn't then multiply up for each population. It's a sum since they are independent groups. Even then, the probability for each subsequent group will be vastly smaller as the population size decreases.
Finally, the only population at risk is users that chose such blindingly obvious passwords that they ended up in a large password population pool. These would be the same accounts that would be revealed almost immediately anyway since the obvious passwords is where any brute-forcing tool worth its salt (pun intended) would guess first.
I agree with you, in particular about the "password storage" stuff. OWASP has a particularly poor history of engaging with cryptography, and, because (I think) of the personalities involved, tend towards a "teach the controversy" approach anytime they're corrected.
Obviously OWASP can't be perfect. We may see the same security issues over and over but the details of particular problems will be specific to the code-base.
But stuff that is universal should be a lot better. Broken Auth page is so horrendous. They mention Broken-Auth then link to a 404 Session Management page, a white-paper on Session fixation, and a paper on password recovery.
It's bad. Instead of being a wikipedia where people can look up types of vulnerabilities, OWASP should try to have more pseudo-code or real code that developers can reference. They have some of this already but they need more
As one of the original authors of clasp, I agree with the poster. We can do so much better. But it's important to realize these things move slow for a variety of reasons. Patches welcome.
Authentication_Cheat_Sheet: password rules which only apply to US-ASCII. That means people with non-US keyboards may not even be able to set a password which meet the rules (eg. Θ and ẞ are upper case, but does not count as upper case in their suggested rules).
Password_Storage_Cheat_Sheet: misses the critical step of unicode canonicalisation and encoding before feeding the password into the slow one-way function (which invariably take an octet string, not textual strings). If you fail to do this, your system will spuriously reject correct passwords if the user logs in via different devices or input methods.
Password_Storage_Cheat_Sheet: '32-bit or 64b' length salt is certainly too short to be called cryptographically strong. Particularly, a 32-bit salt is not enough to avoid leaking password equality with good probability between users once you get past the birthday bound.
Password_Storage_Cheat_Sheet: implies that the caller is responsible for prepending the salt to the credential before inputting it into the slow one-way function. That's not how PBKDF2, scrypt or bcrypt work.