I too have seen the articles proclaiming that devs should stop implementing auth themselves and instead use some Auth as a service product. After all Authentication is hard to get right, right? Actually the basics are really easy. Using a short-term JWT and bcrypt to salt passwords is like maybe 10 lines of code. The thing is, most of the truly hard Auth problems I've had to solve in my career took a highly custom approach anyway.
I suspect most people use products like Auth0, not because they are trying to solve those hard problems, but more likely, because it's easy to set up and comforting to hand off that segment to a company that says they are battle-tested and secure, etc.
I personally like writing my own Auth because I can cater to my personal level of paranoia and learn more. Maybe I'm weird in that way, that I like to consider questions like, how short-lived should my token be? Should I issue a new token on every request? Local Storage vs cookie? Should I use a one-time token? Should I store the JWT in memory only on the client? How would I invalidate a user's token at will?
There's lots of interesting questions beyond the basics for sure and I can see how that might be intimidating. Most projects only need the basics though. And if you need more, it might be easier to extend your own implementation rather than some use-case not catered to by a library or service.
At the very least, I think it's wise for dev's to experience writing their own Auth a few times to get a grasp of some the challenges and better understand what those services might be doing for them.
> Using a short-term JWT and bcrypt to salt passwords is like maybe 10 lines of code.
You are presenting an extremely limited scope.
What you get from using an authentication provider is things like verifying email addresses, OAuth where you just have to "enable" Google, Github, Apple login possibilities. Implementing all these flows yourself is a lot more than 10 lines of code.
Then you also have to do password reset flows, resent email confirmation emails etc.
Yes, that's why I said "the basics are really easy"? This was not meant to imply 10 lines of code handle the complete user experience workflow around auth. In saying that I have implemented oauth for multiple providers in not too much more lines of code. Maybe like 100.
Devs should never default to using JWTs IMO, I think that is leading folks to delegate auth to providers as it overly complicates the process. There's nothing wrong with stateful API keys/session tokens, and I would wager the majority of web services still use them.
Jwt is a much easier term to google for information than the api key/session solution you mention. I actually default to jwt just because I don’t have a clear picture of what the alternative is called.
I agree, stateful sessions can handle a number of problems and use-cases more gracefully. I work with a lot of api's the service multiple clients though and I do see a use-case for stateless as well.
You didn't mention MFA. You didn't mention SSO. You didn't mention self-serve reset flows. Your approach doesn't work well with mobile apps. These are not rare, exotic things in 2022.
Yeah, that's why I prefaced with "the basics of auth are actually pretty easy". Rather than trying to opt-in to those other auth flows before you need them, I'd suggest starting with the basics and extend as needed. Although the MFA I've implemented was actually pretty easy too.
Personal story: Someone in the dark past of my company decided rolling their own auth was fine. Our platform is mostly Django so there's a lot of pieces in place OOTB so they thought it would be easy and it probably was. At some point, whatever this auth system was used for just fell off the websites. Some new login features were added back and they used hosted platform for passwords. Enter me, years later, no idea that auth every existed. Eventually we noticed a database table with thousands of users whose passwords were hashed with a now broken algorithm. They were probably sitting there for years and had copies, backups and snapshots wherever.
I don't quite follow the moral here. Was the Dev who rolled their own auth back when MD5 was the norm supposed to time travel forward to a time when cloud services like Auth and Cognito existed? Or does the fault lie with whoever replaced his auth with cloud services and forgot to drop the passwords table?
Most of the security nightmares I've heard of result from laziness or lack of care or someone not seeing a job through to the end. All of the encryption alogos we use will be easily brute-forced at some point in the future. Is it worse to store and back up those hashed passwords locally, rather than trusting Okta to do it for you (which incidentally was breached while your old MD5 hashes probably weren't)?
Auth providers have existed for a long time. And the point is that when you pay a vendor who does nothing but auth, they will be able to mitigate all the changes in the industry and (if you put it in your contract) be liable for damages from a breach. The also have up to date controls in place for protecting user data and likely go through external auditing. Give our org's limited resources we don't have audit trails on a lot of things and the reality that any number of past employees could have exfiltrated data and we'd have absolutely no way of knowing. Okta had a small breach which they disclosed that didn't seem to leak passwords, and if they did I'm sure they have controls in place to expire anything that may have been compromised. Hashing passwords is easy. Mitigating an attack is really, really, really hard.
Mitigating an attack is really, really, really hard.
How hard is it? What are the challenges? Could you be more specific? Cause I keep seeing the blog and article headlines telling me how hard Auth is but a lot of them don't go into details. That's the kind of mystification I'm having an issue with.
I'm not against using service providers for CAPTCHA or CDN's like Cloudfront for DDOS/brute-force prevention. Refusing users the ability to register an account with breached credentials sounds like a great idea though I've implemented something like this myself already.
I'd wager that aside from what anyone put in their contracts with OKTA, that they didn't pay out a dime for the 25 minute Breach.
Do I want to just pass the buck for the sake of saying, I'm not personally liable or being able to deflect and say, "look this happened to a lot of other companies too". That's part of the fear that these blog posts and articles try to instill with absolutes of "never ever do this..." or "this is way too hard for you..."
I'd argue I'm less of an attack vector than Okta for most of my use-cases. And I think that's part of the issue here, that a lot of us are talking different scenarios and perspectives. If I was an IT head or sysadmin, yeah, I'd probably use something like Okta or onelogin. For a web app with an admin panel for a restaurant taking reservations, Yeah, I'd roll my own and own the users table with passwords as well. For an enterprise-level web app, I might consider compromising and use something like AWS Cognito. But for all my own personal projects, a few paid contractor jobs, and even a few large-scale apps, I have rolled my own auth and preferred it largely to alternative like firebase, and auth0. I'm starting to get the picture that I'm in the minority there. If you don't care about auth and don't want the liability, sure go with a provider. I did mention in my original comment that maybe I was weird in that I liked going down the rabbit hole on auth concerns.
There's an old adage: If you think auth is easy to write, you probably don't know auth.
I obviously don't know how much it compares to your specific case, but as general advice it's often good advice when phrased less antagonistically: don't roll it yourself because the rabbit hole is deeper than you expect.
On that front, you may think you are paranoid enough, but keep in mind that the service products sometimes have a multiplier effect in play: N developers of various levels of paranoia, M people in various ops/InfoSec/auditing roles, X committees/layers designed to minimize corporate security risk.
If auth isn't your job and isn't your core competency, many times it is worth asking if following the rabbit down the hole is worth the time you could save outsourcing to a multiplier effect of a team that focuses 100% on those tasks. As with most "buy versus build" discussions, it's not just about "comfort" but leveraging those multiplier effects. Can you pay someone else for the years of experience and "rabbit hole wandering" and multiple layers of security audits? Auth is a "must have" feature for most products but in and of itself is mostly a security liability and mostly pure overhead: no customer is paying you directly for your auth features, auth is the annoying hurdle to features that they do care about and will pay you for. All that time spent on auth and auth security audits rarely appear to provide good returns on investment for your product and your customers. Time spent on auth is time wasted from user-facing features.
I've never head that adage. I did hear one from Miles Davis that says, "If you have to ask, you'll never know." I ask a lot of questions though so I never paid much attention to old adages I guess.
If auth isn't your job and isn't your core competency
I'd argue, it sort of is a core part of any web dev's job and any who neglect diving into it do so at their own risk. Maybe I come from a different time or place than what's the contemporary norm, though I'm okay with that. The main issue I have is with company-sponsored blogs and comments that will just blanket state you should never roll your own auth (because you can pay us instead, oh and btw, we rolled our own auth so you don't have to).
I can understand skepticism with company-sponsored blogs. For what it is worth, I don't have a product to sell here in this space and make no commissions.
As a user, I subscribe to haveibeenpwned alerts on my main email addresses and have been for multiple years. I think there are fewer humbling "[hackers leaking my users table] won't happen to me" reminders than that. It happens to everyone. The ones listed in haveibeenpwned are just the tip of the iceberg and the ones we know about and have been confirmed/verified.
As an enterprise developer, I've been a part of some of the worst security audits imaginable. Thankfully those were all preventative industry requirements, and I'd hate to have to deal with "the real thing". Those audits always come with a reminder: the most secure data is data that you don't have, anything else it doesn't matter how well you think it is encrypted at rest is a liability.
As a user I never want to have to setup another password ever again. I realize that's a pipe dream and we are still at least a few standards away from making that dream a reality, but I believe in it and I've championed it when I've thought standards got close. (I was a fan of the original, now dead OpenID [but not the Zombie OIDC wearing its skin]. I thought Mozilla's Persona/BrowserID was brilliant and killed too early in its infancy. I'm hopeful for Webauthn.)
As a developer I never want to own another database table with passwords stored in it. It's low hanging fruit that every attacker is looking for. I don't care at this point how those passwords are salted, hashed, encrypted, mashed, or other waffle house preparation keywords. Attackers still try to pull the records no matter how safe you think they are. Not having them at all is just entirely easier than worrying about having them.
I don't entirely care how I get rid of the password table. I've paid for solutions. I've used passwordless tools like "magic links". (To use those right you often need all the hard lessons of JWTs and other token types and keeping those secure. It's not necessarily "easier" or "harder", but a lot of knowing how to do it right transfers.)
I just believe pretty strongly at this point that the best password table is one that you don't have and don't have to manage. No matter how easy it seems to "just bcrypt it".
There's a lot of food for thought there and I can actually get on board with a lot of what you said. I enjoyed your perspective at the very least. I too have been a part of some tough audits, fedRAMP, etc. I've worked on apps that regularly get pen-tested. XSS is real, though more often I've seen ITOR issues. No need to worry about being hacked when you're giving it away for free. And then there were at least four OWASP courses I've been asked to do for work.
I'm not quite to the point of never wanting to own another user table with a password column. Maybe I will be someday. Yes, it's a liability but I still think my table (which lives on RDS or Aurora btw) is less of a target than Okta's users tables, which were breached earlier this year. BTW, Auth0 was recently acquired by Okta.
But yeah, if I was spinning up a large-scale, enterprise-level web app, I'd probably tell my stakeholders we were going with AWS Cognito or something, just to offset liability. Never mind, I'm still going to have to have a users table, and if I'm using Cognito for authorization (group-level authorizers) then god help me (check out their limits on those custom fields too). And btw, they use JWT, and their front-end solution Amplify still uses local storage to hold credentials. And it's going to be even more difficult to set up the basic auth flow than if I rolled it myself, but who cares, as soon as the execs hear AWS solution or cloud-native, they nod and feel assured. After all, the liability is now somewhere else as you said. Never mind there's a host of missteps and other ways you could still mess it up on your side of the fence there.
For all of my personal projects, a few smaller paid side-projects, and at least two enterprise-level products, I've rolled my own and owned the DB and it's felt much better to me than the alternatives I've had to face when I forced into some auth as a service solution. I've done auth0 twice (never again). Firebase once. Cognito at least 4 times. I still prefer to roll my own. Maybe I am a masochist. Maybe I've been lucky. Maybe I just go against the grain and do things my own way, providing less of a common attack vector.
The number of devs I've run across who don't even understand the basics of authentication has surprised me. Yes, it's seemingly easy in concept as you say, to "just bcrypt it" but it astounds the number of devs I've come across or worked with who didn't even understand what that was or what it looked like. Then we have the "company-sponsored blogs" and the echo chamber of people throwing down absolutes of "whatever you do, never roll your own", "you don't understand how hard Auth is", etc...
But I'm coming to accept that maybe I'm in the minority on this topic so I'll stop ranting for now. I suspect that a lot of us on this thread are speculating or speaking from different scenarios and experiences, like a sysadmin or IT head, vs legacy enterprise apps, vs mobile, front-end specific SPA's with stateless auth, vs server-rendered stateful, session-based web apps, etc...
In any case I appreciate you taking the time to map out your thoughts and perspective in detail and I think it's probably broadened my own thoughts on the subject.
Recommending people that don't know how to/if to do it (that's why they are asking) to write their own auth, trivializing the problem and in the same comment recommending using JWT terrifies me quite a lot.
There's nothing wrong with JWT as a token spec. I prefer stateful sessions as well but but a stateless token also serves specific use-cases. It has also become a standard practice for decouple front-ends that live on different domains, etc.
> I too have seen the articles proclaiming that devs should stop implementing auth themselves
> and instead use some Auth as a service product.
It's a moved goal post which sneakily tries to position itself on top of the old adage "never roll your own crypto". Problem is, that adage is really about security whereas "don't roll your own auth" is about not wanting to do a mostly boring part of your job. I'd take my hat off if they'd admit it, but it's much easier to get budget with potential security scares compared to "we don't wanna".
Having your own password table even feels archaic today. From what I've seen, even simple internal apps are doing SSO with a directory provider (Google, AD/ADB2C, Okta, etc.)
I mean, it's stored in a DB somewhere, even if it's not your own. I would argue your own DB probably has less of an attack vector than major providers.
Existing crypto libs like bcrypt, which I mentioned? Or are you talking about something more? The credential stuffing mention you make makes me think you must mean something more. Does Auth0 by default advise users that they are using breached emails/passwords? If that is default behavior then I agree it does add a shining point to their feature-set.
captcha integration is super easy and doesn't require a full auth integration service. Detecting a brute-force maybe less so maybe, though I get a lot of DDOS pretection for free with AWS & cloudfront. But I was more interested in the idea of letting users know they were trying to use a breached email/password on register or login. I've got that on bitwarden (password manager), though now days I just auto-generate any password.
Not implementing it yourself does not imply that you should use auth-as-a-service. There are plenty of auth libraries out there for your programming language of choice. Take a look at their source code and see how much stuff is in there to get a sense of why you should not implement your own.
Oh yeah, for sure, there are a really good libs. Someone mentioned passport. I've personally used Devise/warden in Rails. Though we had to monkey patch/extend it so much that 6 years later it was almost a different beast entirely.
But I'd still argue it's worth knowing what those libs are doing for you and you can only really appreciate it once you've rolled your own. I've done so on many a personal project where the risk was low. But yeah honestly, if I was rolling a large-scale enterprise app today, I'd probably utilize AWS Cognito. But this is also not for the faint of heart IMO.
In your “should I..”s, you mention implementing multiple insecure approaches. If you’re implementing sessions tied to localstorage, your application has a medium severity vulnerability. If you’re improperly invalidating, that’s yet another medium severity vulnerability.
The problem is, you’ll slip up here and there. The experts recommend not rolling your own for a reason; little mistakes and gotchas like the details in your post can lead to significant compromise.
I mean, yeah, sure. Or for a number of other use-cases. I'd argue the best way is for the client to keep the token in memory over local storage. Of course if you have a tightly coupled app, session-based secure cookie is the best.
What I mean by tightly coupled, is if you can have your Frontend and Backend on the same domain, then yes by all means an httpOnly cookie is ideal. Now days there are a lot of SPA's that don't do that, or that have multiple api services across domains, or you are only writing a front-end and using some backend as a service like AWS amplify or numerous other scenarios that have led us to the concept of stateless authentication. You can still use a cookie in those situations, albeit without the use of httpOnly, which means they are going to be just as vulnerable to XSS as local storage. This is why I mentioned keeping a token in memory (this can be made even more secure with web workers). But since this doesn't persist across reloads, it's not an ideal solution for most leading towards utilizing strategies using refresh tokens/one-time use tokens, etc.
I suspect most people use products like Auth0, not because they are trying to solve those hard problems, but more likely, because it's easy to set up and comforting to hand off that segment to a company that says they are battle-tested and secure, etc.
I personally like writing my own Auth because I can cater to my personal level of paranoia and learn more. Maybe I'm weird in that way, that I like to consider questions like, how short-lived should my token be? Should I issue a new token on every request? Local Storage vs cookie? Should I use a one-time token? Should I store the JWT in memory only on the client? How would I invalidate a user's token at will?
There's lots of interesting questions beyond the basics for sure and I can see how that might be intimidating. Most projects only need the basics though. And if you need more, it might be easier to extend your own implementation rather than some use-case not catered to by a library or service.
At the very least, I think it's wise for dev's to experience writing their own Auth a few times to get a grasp of some the challenges and better understand what those services might be doing for them.