What's worse is that they've known that it's an embarrassing security flaw for years, but they apparently thought other things were more important. Not really very comforting.
"Embarassing security flaw"? Come on. We audit several major web applications every week --- things that are far more sensitive than 37s, like banking, trading, and pension --- and we see sites store plaintext passwords all the time. I've never seen any vendor's security report write someone up for this. If it did get written up, it'd be "Severity: Low", along with "enabling weak SSL ciphers" and "exposed server version information".
I'm not trying to be an apologist here, but "someone is wrong on the Internet" if they are saying that storing passwords badly is a major vulnerability.
Is it a big deal to properly store the passwords or isn't it?
> stop working on your social shopping cart calendar application right now: I can’t trust you with my Reddit karma score, let alone my credit card number.
and that's what you said for developers using fast 1-way hashes. So it's embarrassing to use md5+salt, but not embarrassing to use plaintext?
The method we use for storing passwords is either vital or it's not-- I'm currently lost as to how you could respond like the above and have such strong feelings about the matter in other discussions.
Two thoughts I'm asking you to keep in your head simultaneously:
(1) It is very bad to store passwords in plaintext, with a naked SHA1 hash, or with a naked SHA1 hash and a salt of any kind.
(2) As bad as it is to do that, it is not a critical security vulnerability in your application to make that mistake.
The classic web pester example of a "Medium" severity vulnerability: reflected cross-site scripting ("here, follow this link --- hah, now I own your session!"). If you tell me that reflected cross-site scripting is less bad --- or even the same level of bad --- as storing passwords poorly, I'll accuse you of being disingenuous.
See downthread, upthread, and sidethread for examples of me pointing out my stance on safe password storage.
Sorry, man, I kind of can't take "disingenuous" accusations seriously as that was the word on the tip of my tongue when I read your surprising statement earlier in the thread. It doesn't sound like the same person I read before, nor from someone with your (deserved) status as a learned engineer on security matters.
The point you disagreed with was whether this was an embarrassing security flaw:
> "Embarassing security flaw"? Come on.
and that's what I'm calling you on based on our earlier discussions. I'm not making any point about how this compares to other security flaws, I'm just trying to figure out where you really stand on plaintext.
An earlier quote from you as I try to sort this out:
> People get to access password hashes as soon as you mess up a single database query. The idea behind storing safe hashes is to prevent your stupid mistakes from screwing over every one of your users. The stupidest people of all are the ones who assume they aren't going to make stupid mistakes.
So, the question at hand: Is it an embarrassing security flaw to use plaintext passwords for a modern, respected website?
It may be embarassing, and it may be a flaw, but if you sent it to Bugtraq or asked ZDI to pay you for it, you'd get laughed at. I'm not making that up, it's obviously true, so why are you arguing with me about it?
If you contract a team from Matasano to assess your website, and all we come back with is "you store passwords insecurely", we're the ones who will be embarassed. Sorry. Hate on 37s all you want, but in no parallel universe is "weak password storage" a gameover.
OK, so then to be sure I'm understanding, let me walk through how Django (in the default auth backend -- this is customizable if you want to plug in another scheme, as people often do) does it:
1. The "password" field in the DB holds a value of the form "algorithm_name$nonce$hash".
2. "algorithm_name", by default, is SHA1, but the backend lets you also use in MD5 (for backwards compatibility) or Unix crypt if you're running on a platform which has it.
3. "nonce" is the first five digits of the hex digest of a pair of floats pulled (at the time the password is set) from Python's 'random.random()' (which, of course, is a PRNG, not a truly random source of numbers).
4. "hash" is the result of applying the algorithm to the combination of the plain-text password and the nonce.
It's a bit confusingly set up internally because the code refers to the nonce as "salt" even though it's different (within the limits of Python's PRNG) for each stored password. And there's some fun legacy stuff in there that transparently switches naked MD5 hashes over to the algo/nonce/hash scheme when possible, which makes even more fun.
As best I can tell, this is the sort of scheme you're recommending, or at least that you're not outright bashing. Is that correct, or have I missed something?
We've had requests to open up the variety of hashing algorithms supported in the default backend, and AFAIK the jury's still kinda out on that one due to various compatibility issues with existing DBs and older Pythons, and the ease with which alternate authentication schemes can be plugged in.
The Django scheme is what tptacek refers to as a "naked sha1 plus salt" and recommends against, because it is computably feasible to brute force it due to sha1 being a fast algorithm. His recommendation is to use a more computationally expensive hash such as bcrypt (based on blowfish) or possibly a few thousand iterations of sha1.
I guess this is where I'm getting confused. In some places I see comments which seem to say that "naked hash plus salt" is bad when the salt is the same every time, but possibly not when the salt is different every time. And, of course, in other places I see comments saying that this isn't any good either. And then stuff which has nothing to do with salts or nonces but solely looks at the algorithm (which I don't understand, because even the most up-to-date stuff on attacking popular hashes still seems to be orders of magnitude behind the efficiency of rainbow tables, and rainbow tables don't do as well against a salt that changes with every password).
And, well, this is where I just kinda go off the rails, because it seems as if there's nothing that's ever been conceived by man that doesn't have an article by a security expert saying "what are you, some kind of idiot? Why would anybody use that?"
I don't believe that this topic is, in practice, hard to understand.
You do face the challenge of sorting through all the noise of proposals and retorts --- "what if I use a 64 bit salt?", "what if I use SHA-256 instead of MD5", "what if I use AES to encrypt the passwords", "what if I run SHA1 in Javascript and just exchange the hashes", etc, etc, etc. None of this stuff matters.
The core issue is: your password hash should be cryptographically slow. It should resist account-at-a-time brute force attacks, which are the predominant method used to crack accounts. You can safely ignore the rest of the sideshows.
And finally, I have to say that just because a topic seems annoyingly complex doesn't mean it morally has to be simpler. If security experts are telling you you're an idiot for doing X, by all means get pissed that they called you an idiot, but also stop doing X.
"And finally, I have to say that just because a topic seems annoyingly complex doesn't mean it morally has to be simpler."
I'm OK with complexity. What I'm not OK with is the impression I get that there simply is no option that won't get shouted down. "Making the right choice is complex" is fine when compared to "there is no right choice".
Nobody will shout you down for iterating SHA1 or for using one of the bcrypt/scrypt/PBKDF.
In actual fact, I won't even shout at you for using salted single-pass SHA1, as long as you don't brag about it, or talk about how your "salt" construction improves security.
Well, in an ideal world I'd be in favor of using something better than what we currently have as the default, but for now we have to strike a compromise -- SHA1 is about the "best" hash that's ubiquitously available on the platforms and Python versions we support. Down the road when we've deprecated a couple older Python releases and can rely on things like hashlib being available, it'll be time to talk about improving the default setup.
And, in the context of common use cases, I think it's probably an OK compromise -- I doubt somebody's going to devote the resources to plucking the password for a random blog, for example. Once you need tougher security, you can plug in whatever scheme you like for storing credentials and authenticating users which, anecdotally, most people seem to be doing when warranted (there are quite a few third-party auth backends floating around). And, really, if I were going to brag about anything, it'd probably be the ease with which you can do that -- one class with two methods, a settings tweak and (depending on your credential scheme) possibly a short form class that knows what the incoming data will look like, and you're done. Of course, I think we could make it even easier, so I won't be bragging about it just yet :)
SHA-512 and Whirlpool are still designed to be very fast. The point is, you want to slow down the password hashing operation. So SHA-512, SHA-256, and SHA-1 are all fine, as long as you iterate them several thousand times.
I worked on a bank's web application and I recall our debating this very thing vigorously. We did have security audits, and if we did something like this we would expect the subject to come up in an audit.
The reason we wouldn't have considered this to be a "major" security vulnerability is that compromising our database would already be a major catastrophe, so losing a few million people's passwords would be the least of our worries.
So for a bank, it's like the old joke about not needing to run faster than a bear, just faster than your buddy.
However, my guess is that 37Signals is a little different. We use campfire, but we do not discuss certain thing son it specifically because it is hosted on a 3rd party server. If 37Signals' database was compromised, we would be distressed to have our private conversations exposed, but we would be mortified if one of our developers' passwords was compromised and that became an attack vector into more sensitive information.
So I am suggesting it comes down to relative vulnerability. For a bank, a compromised database is a catastrophe whether it contains plaintext passwords or not. For an online chat application, a compromised database is not a catastrophe, but compromising user passwords might tilt the balance from inconvenient to embarrassing.
The slight fallacy there is assuming that it takes a full database compromise to extract the data.
One of the things I have found is that often data (especially high-use data like passwords & usernames) is able to be snuck out through much more creative routes - storing in plaintext means extraction is much more worth the effort :) no waiting around to crack the data instant use.
If I sneak the passwords out, can't I also sneak customer personal data like bank balances, SSNs, and so forth? Most banks maintain more than enough information on a customer to perform an identity theft and then some major fun begins such as misappropriating the title to real estate.
I don't disagree it can be done, just pointing out that if we assume a compromised employee and a vector for removing the passwords, encrypting/hashing the passwords is like locking the side door of the stable.
But hashing the other data isnt practical - so it's the security vs. access tradeoff. You can hash passwords without much inconvenience.
Also with individual pieces of those other examples of data you mention it is not possible to gain access to the rest. Whereas with a password it IS. Damage limitation.
Finally the password and username (or email) are "front of the line". They HAVE to be accessible to an unauthorised client simply so said client can be authorised. Whereas the other data should require authorisation to access.
The fact that this is a common mistake doesn't make it any less serious. I don't do security audits, but if I did I would certainly point out this sort of mistake -- because even if it doesn't affect this site's security very much, it certainly has an impact on their customers' security thanks to the reuse of passwords.
I'm going to have a hard time arguing that it isn't a mistake given my track record on the site.
My only argument here is that, mistake or not --- it's definitely a mistake --- it's not a "serious security vulnerability", at least not as the industry understands them.
From a theoretical security perspective, it's a terrible idea to store passwords in plaintext, but again there's a difference between theory and practice as you've stated. If your auditing can reduce/eliminate unauthorized out-of-band database access, then in practice storing plaintext passwords isn't a huge problem.
However, there's another practical aspect to consider: the password security of users. A perfect example from a month ago is a Christian dating website called Singles.org. The 4chan crowd discovered that if you replaced your user ID on the "profile edit" page with another user ID, you could edit another person's profile. The page also helpfully displayed the registered e-mail address and password, in plaintext.
Now, this is terrible programming on multiple levels, so you might want to discount my anecdote, but the attack vector isn't my main point.
The 4chan crowd, ravenous beasts that they were, attempted to use the Singles.org password to log into the victim's e-mail account. And their Facebook account. And their PayPal account. You can imagine the chaos they caused.
The user base had terrible password security, so easily around 20% of the accounts had the same password across multiple websites. I don't think it's an uncommon statistic: sadly, even I as a security-conscious person have used the same password more than once.
Our efforts should be to secure an application against all attack vectors, but hashing passwords is a veritable last defense to prevent an attack from spreading on to other websites.
Regardless of how an application is exploited, it's a safe assumption that the users' poor password security could spread the attack further. I would thus claim that plaintext passwords be at least a "medium" security issue.
Being able to get access to plaintext passwords is obviously a terrible security flaw. But would it be much less bad if an attacker got access to a lot of hashed passwords? A dictionary attack will get you many of the plaintexts in a reasonable timespan. Possibly users with dictionary words as passwords are also more likely to share them across sites; certainly the intersection is nonempty.
People with strong passwords might remain safe, but not necessarily. A bruteforce is much more worthwhile when you're trying to crack thousands of passwords at a time. If there's no salting and the hash algorithm is widely used, you may not even need to worry about getting exactly the original rather than something which happens to collide with it.
Obviously hashing is still better than no hashing, especially since there will be cases where only a few passwords are leaked. But if a database containing passwords is stolen, it may be reasonable to assume all those passwords are compromised, even if they were hashed.
If the password hashing scheme is secure (bcrypt, as tptacek repeatedly recommends) dictionary attacks become non-trivial and slow. It's still possible to brute force the database, of course, but the time required makes the attack less profitable.
I would thus argue that a secure hashing schema is significantly better. This is assuming the programmer uses bcrypt in the first place, which is a questionable assumption: even popular projects like phpBB use MD5 for password hashing. (The fools.)
If MD5 is used then I would wholeheartedly agree with you: I've used http://milw0rm.com 's MD5 hash breaking service plenty of times to know why I shouldn't use that algorithm in my own code. (One of my old passwords is in the database, but I'm not telling you which one.)
To me it's a "serious security vulnerability" only in the sense that if they were that lackadaisical with their password storage, where else have they made poor decisions with regards to security?
I think you'll agree that security is not a single thing you do, but a series of things that makes getting at sensitive data more difficult. To store all passwords as plaintext (including passwords for administrators, I would assume) surely must be considered a serious oversight at the least.
I feel as though maybe we're missing some nuance in your position on this.
I didn't say it was a major vulnerability; I said it was embarrassing. Given 37signal's stature in the web community, the ease with which this problem can be fixed, the fact that they said they would fix it years ago and haven't, and the fact that this is a basic problem with a well-known solution, I think this is embarrassing for them.
Really? We're going to downmod the most active security professional on HN? A guy who's built a 25+ person company on the basis of analyzing web applications for security? FFS, did we learn nothing from yesterday's expert kerfuffle?
If Thomas says it's not really a big deal in practice, then it's not really a big deal in practice.
Note that I might not be "the most active security professional on HN", but am certainly "the security professional most active on HN". There's a big difference. ;)
Note: I don't know who this "kubrick" person is, because he won't tell us, but "CISO of a financial institution or not", he's wrong about the OWASP Top 10.
Neither "broken authentication" nor "poor encryption" "top vulnerabilities" discuss password encryption. "Storing plaintext passwords" is not an OWASP Top 10 flaw.
"A7 Broken Authentication and Session Management" is a proxy for session fixation, and the catch-all category for things like "cookies that store the bit that tells the app whether you're the admin. It mentions passwords repeatedly without ever recommending they be encrypted; in fact, by talking about "Question and Answer" features, it tacitly acknowledges the practice.
"A8 Insecure Cryptographic Storage" means "don't invent your own crypto", and now also "don't use MD5". It mentions passwords zero times; the only thing it says need to be encrypted are credit cards. This makes sense, because if it said anything else, no real-world web app would pass muster.
Having said all that, I wouldn't put much stock in the OWASP Top 10 either. It's the most popular list of web flaws, but it isn't the most complete, nor is it the most coherent. It's not like there's science behind it.
On the contrary, it's a big deal to me because it's my password and I judge that it's being stored in an insecure manner. I really appreciate Thomas's expertise, but since my data is at risk as a customer, my opinion is really the one that counts.
I'm actually bizarrely fixated on this topic; go to "searchyc.com" and search for "bcrypt" and follow the threads. It is, apparently, all I ever talk about.
I'm not defending the feature. I'm just saying:
* Lots of companies have this feature.
* Among them are FedEx and several banks.
* If you wrote a security report for Basecamp and included this problem at all, it would be "Severity: Low". I'd write it up. But I wouldn't say "Severity: Critical", because I would get my ass kicked by my clients and partners.
The rest of this debate, have at it. I agree. BURN THE WITCHES.
Well also as a security professional (also called Thomas as it happens :P small world) I would point it out as at least a medium security flaw.
> because I would get my ass kicked by my clients and partners.
So your reasoning is nothing to do with security - it is to do with saying what the customer wants to hear (low severity stuff can be ignored, right?).
I have found the opposite in the past: encrypting passwords is, I agree, a common issue and one that is also usually very simple to fix. You can present that whole fix to the client and it can probably be implemented in a few weeks/months - it's useful because it is not something that will send them screaming to the hills but it is something so they feel they are getting value for money :) (plus you will likely get follow up work - if not to fix the problem then to audit the fixes!). Win win.
At the end of the day storing in plain text means it only takes on slip to release all your users passwords into the wild. It IS a big security flaw. If I walk away from an audit w/o flagging the password encryption as something to be addressed fairly soon and then the passwords are stolen my ass is properly on the line (probably more than any other issue).
And for the record I am talking as big if not bigger institutions than Fedex and US banks.
My reasoning is that usually we call out things that can actually allow an attacker to compromise the site, and don't spend as much time on the million things that might make an attacker's life easier after that compromise has occurred.
The rest of your argument is a moot point, because I'm not asserting that passwords should be stored plaintext.
It's not really moot because I never assumed you were asserting that.
But you did just say that you would either not mention or flag as low priority (and the suggestion is grudgingly) a plain text passwords issue because "I would get my ass kicked by my clients and partners". I'll be honest - I wouldn't hire you if I had seen that written publicly like that :(
I sometimes think that one of the major failings in our industry is that we toe the corporate line and never stop to think like a cracker. That's why I have my job - because I do. Some people see plain text passwords as a "making life easier" issue, whereas for me it is a major weakness at the core of the security chain. Throw all you like in the way but, at the end of that day, the passwords are there in clear text waiting for me to find it. It doesn't matter how many different ways I can or cant compromise a site: all I need to do is do it once...
And anyway, my point was less about this specific issue than about how I think you address it wrong. The hashed passwords issue is one you can dress up for a client so they think they get value for money. If they get nothing except a green tick on the "critical errors" page then you leave them with the feeling that they are missing something. Hashing passwords should be fixed - and we can get them to fix it. And at the same time they get something meaningful from their audit :)
* The probability of your servers getting hacked/stolen is reasonably low as long as you take other precautions.
For most people, I'd say it's a lot of extra work, for no real pay-off - apart from making your users feel a bit more warm and fuzzy.
A better solution would probably just be a disclaimer on websites saying "Please don't use the same password you use for online banking here, as that would be silly".
Famous last words: "If it was really big, it would have been exploited already."
Never mind industry best practices or the cautionary incidents at other companies -- this "it's not big until it bites us" attitude can be used to deprioritize fixing any risky behavior until it's too late.
There are low-stakes situations where this attitude is a virtue -- where the damage from an exploit is minor and quickly reversible. For famous and business-class internet services, the stakes are high.
