Hacker News Reader

Comment by jcranmer | Hacker News Reader

In my experience with OAuth, one of the principle issues is that it's less a protocol and more a skeleton of a protocol. Actually go to the core RFCs for OAuth, and you realize that basically everything is implementation-dependent. You somehow register with the provider as a client (out of band and completely implementation-dependent).

Then you ask them to log you in by sending a request to an implementation-dependent webpage, with implementation-dependent parameters (which ones are required are--you guessed it--implementation-dependent), telling them to redirect you to your page when you're done. Well, actually, that assumes you're building a website. If you're a desktop application, you'll do something else. If you're on an embedded system that can't open up a webpage, there's another option. And there's yet more flows. Which ones are supported by the provider? You guessed it, it's all implementation-dependent. Oh, and maybe you need to refresh tokens to login the future. When? If? You guessed it, implementation-dependent!

It makes writing a generic OAuth client library hard because there's basically no commonality there. Really, it makes me long for Kerberos as a much saner SSO system than OAuth.

View on HN

> In my experience with OAuth, one of the principle issues is that it's less a protocol and more a skeleton of a protocol

OAuth supports lots of different scenarios. Many people when they say “OAuth” they are only thinking of one or two of those scenarios, and ignoring all the others which aren’t relevant to them personally-but may be relevant to someone else

I worked on a system where we had a micro-service which had a token from OAuth server 1, and it needed to exchange it for a token from OAuth server 2, and we needed some policies to decide whether that token exchange was allowed or not. And that’s totally a use case the OAuth RFCs support (there is even an RFC specifically on token exchange), but a person wanting to add a “login with Google” button to their website isn’t interested in anything remotely like that.

> Which ones are supported by the provider? You guessed it, it's all implementation-dependent.

Security needs vary widely from system to system. So they defined a protocol which supports many different scenarios. But if your application only needs two of them, why implement the other N? OTOH, for someone who actually needs one of those other scenarios, having it standardised makes their life easier. You can’t expect the protocol to tell you which scenario you have, that’s inherently “implementation-dependent”, to the point that calling it that is getting tautologous

> > In my experience with OAuth, one of the principle issues is that it's less a protocol and more a skeleton of a protocol

> OAuth supports lots of different scenarios.

But it's literally only the skeleton of a protocol ("framework" of a protocol, as the actual spec puts it). That important "Then later on you can take the token and say 'Hey google, somebody gave me this token, who is it?'" part is totally unspecified by OAuth. How do you validate tokens? (Either on the resource-server end, or on the client-end to know if trying to talk to a resource-server is a waste of time, or to know if your post-redirect-uri is getting spoofed):

- OAuth: implementation defined

- most implementations: use our library, which validates the token using undocumented logic

- OIDC: there's a user-info endpoint you can call to, but lots of OAuth authorization servers don't fully implement OIDC, and even if they do, that's a lot of extra round trips

- RFC7523: The token is a JWT, validate the signature and claims... but everyone issues RFC6750 "bearer" tokens, not RFC7523 "urn:ietf:params:oauth:grant-type:jwt-bearer" tokens. But if you just close your eyes and pretend that your 6750 opaque token is actually a 7523 JWT and parse+validate it as such, that'll work most of the time.

So yeah, OAuth is hard because there's no great generic library, because a core part of it is implementation-defined, so you have to either do it yourself or use a specific implementation's library, and none of those libraries work quite the same.

(In the above, and for 99% of folks, "OAuth" = RFC6749+RFC6750)

In all honesty, though, "I have to look at the provider's documentation to determine like 4 URLs to add to a configuration" is not that wild of a thing? How much friction is too much friction?

> How do you validate tokens? (Either on the resource-server end, or on the client-end to know if trying to talk to a resource-server is a waste of time, or to know if your post-redirect-uri is getting spoofed):

At the beginning of the "standard" OAuth flow you can pass in client state. So you generate a signed nonce. Remember, you are starting this flow, so you mix in some user data (avoids CSRF problems), and you are pointing to some HTTPS site.

This is not 100% perfect compared to "remote server has some signing" but it does allow purely local verification of tokens when they are first received, as you can stick stuff into the "state" request parameter (as anyone should be anyways to avoid trickery).

> In all honesty, though, "I have to look at the provider's documentation to determine like 4 URLs to add to a configuration" is not that wild of a thing? How much friction is too much friction?

If it's something dumb/simple like grabbing a JWKS URL and an "issuer" URL, and then doing JWT validation, sure. Maybe you need to make a weird bespoke request to some other endpoint. Maybe you need to implement some uncommon signature verification. It could be anything.

> At the beginning of the "standard" OAuth flow you can pass in client state. So you generate a signed nonce. Remember, you are starting this flow, so you mix in some user data (avoids CSRF problems), and you are pointing to some HTTPS site.

So you redirect the user's browser to the IDP, with the nonce in the URL. The user grabs that nonce from the URL bar, then manually navigates to your redirection-endpoint, putting that nonce and their own fabricated token in the URL.

...what?

The server generates the auth code and redirects the user agent to your callback. You exchange that code with the IDP (over HTTPS which yeah that's its own nest of wormy trust) to get back a token. They can't inject a token because you don't get the token from them, just the one time code. If it's opaque you introspect it to validate or you just validate the JWT signature after pulling the keys from the JWKS endpoint. Introspection is standardized and an RFC. The state param is just a fucking session identifier.

All these URLs are defined and provided via the .well-known/openid-configuration endpoint. If your IDP publishes that endpoint correctly, most OAuth2 client libraries Just Work (TM) when pointed at the IDP domain.

Do EITHER of you even use OAuth2 outside of just cargo culting something you found off GitHub?

Oh, thank you for reminding me of the well-known endpoint, was having trouble finding it in the OAuth RFCs but I wonder if it's pulled in elsewhere.

> They can't inject a token because you don't get the token from them, just the one time code.

that's not correct in the most common flow? The most common flow involves the user agent providing the information via a GET, so they can theoretically provide a token.

The reason the state parameter is important is because without it, a malicious actor can make a link that goes directly to your system's "oauth step finalized" step, but with their credentials. (Pedantic attack: service has slack push notifications integration, through OAuth. Attacker creates link like https : //service/connect-slack-finalize?token=token-to-attackers-slack that victim clicks on. Without using state, the service will just take the token and stick it into some slack integration. now attacker's slack is getting messages for the victim's account on the service. Cookies mean that the victim's thing is accepted immediately).

That's something that isn't OAuth2 or your end point is accepting something insane.

Are you talking about the PKCE variant of authorization code flow which is what replaces implicit flows in native apps and SPAs? Because those use code_challenge and code_verifier fields, not the state field. If you're doing all that in the state field with signed nonces you really should move to PKCE.

Using [0] as a reference, I'm talking about Step 3. This is, in my experience, the "normal" way that people are setting up OAuth between 2 services, with a user going through the flow.

[1] includes info on this (see "flawed CSRF protection")

[0]: https://www.digitalocean.com/community/tutorials/an-introduc...

[1] https://portswigger.net/web-security/oauth

Aha! That makes sense! Yes that can be a problem. We exclusively use a single (our own) IdP so it's less important for us. But good to know as some future feature work will actually make this important.

Thanks!

It's to prevent CSRF attacks. The attacker writes their own client and does half of a login on their own end (getting an authorization-code, but not yet exchanging it for an access-token), and then tricks the end-user to navigate to //service/connect-slack-finalize?code=<attackers-code>&state=<whatever>. But with the state parameter, the client can check the state parameter against a session cookie that it set previously, and say "wait a minute, this is the conclusion of a login from a different browser". Depending on what all session-state the client is keeping track of, it may make sense to sign that state-parameter-nonce to avoid having to remember session state server-side; but the simple case would be to just check whether it == a cookie value.

.well-known/openid-configuration is specified as part of OpenID Connect (OIDC) Discovery[1]. OIDC is separate from and on-top-of OAuth2. The OIDC specs come through the OpenID Foundation, not through the IETF (so not RFCs). (Also, while what they specify is super useful, they aren't nearly as well written as RFCs tend to be :) )

[1]: https://openid.net/specs/openid-connect-discovery-1_0.html

> that's not correct in the most common flow?

No, he's right, I was misremembering (assuming "the most common" flow is the "authorization code" flow specified in RFC 6749 §4.1). The user-agent provides a one-time "authorization code" to the client via a GET, and then the client receives that "authorization code" and does its own POST to the IDP to exchange that "authorization code" for the final "access token".

This is me misusing the word "token". Access tokens are gotten via POST, but the one-time code is gotten via GET and, absent usage of things like the state parameter, can easily lead to malicious attacks.

My apologies, it's been a while and I was forgetting about the authorization-code exchange step. So yes, for the most part "client"s can treat the bearer token as opaque. But "resource server"s absolutely cannot.

> If it's opaque you introspect it to validate or you just validate the JWT signature after pulling the keys from the JWKS endpoint.

If it's a JWT, which it doesn't have to be. OIDC allows the token to either be an RFC 6750 opaque token, or an RFC 7523 JWT, and overwhelmingly implementations use a "bearer" (6750) token. But, most of the time it's a JWT, so as I said, closing your eyes and pretending it's a 7523 token works, and so then you can just pull the keys from the JWKS endpoint.

> Introspection is standardized and an RFC.

In my experience, it is exceedingly rare for an IDP to implement RFC 7662 token introspection

> All these URLs are defined and provided via the .well-known/openid-configuration endpoint.

Yeah, OIDC-discovery is pretty sweet. And if the IDP implements OIDC-discovery, then it probably implements OIDC-core, or at least enough of it that you can use the user-info endpoint, like I mentioned. But I've seen IDPs that don't.

Have you ever used OAuth2 outside of OIDC?

One thing I could imagine here is a signature flow happening from the sender side where they sign the nonce + the token, to avoid forgery. I don't know what the signature validation looks like in that model though (if you request a public key from the provider, are you requesting that at every request flow? If you're not, now you're adding a request to every normal, non-malicious validation flow)

> Maybe you need to implement some uncommon signature verification. It could be anything.

So my feeling on this is that at the very least OAuth tends to be "well it's going to look like this, but there might be some tweaks".

I have set up OAuth verification with multiple services in the past. The social aspects and the business aspects have always been more complicated, but at least "it's OAuth" (much like "it's REST" for APIs) establishes a core understanding very quickly.

I'm open to saying that we should add more standard stuff to the flow, if it's optional then people who use oauth libraries will end up with a lot of this stuff by default. But I think the status quo is pretty nice, all things considered.

My apologies about "to know if your post-redirect-uri is getting spoofed" and then my example exploit; I was misremembering and forgetting the authorization-code step.

If you're only implementing a "client", then yeah, you can probably get away with never caring about inspecting the token. But if you're implementing a "resource server", then you'll need to, and it's all implementation-defined (but overwhelmingly that implementation is "it's a JWT").

Any friction whatsoever is too much friction. By default, any user should be able to use any provider for any service, with no prearrangement between the service and the provider.

Anything less than that predictably leads to the situation we have now where a very small number of very large providers control identity and authentication for way too fucking many things.

Yikes. As someone who is going to need to integrate OAuth using JWT into a Rails app to access an API in the nearish future... I now dread it.

> And that’s totally a use case the OAuth RFCs support (there is even an RFC specifically on token exchange), but a person wanting to add a “login with Google” button to their website isn’t interested in anything remotely like that.

That's because OAuth in the industry has been changed to only talk about authentication (AuthN) and not (or very lightly) authorization (AuthZ).

And, for the better, AuthZ is so use-case specific that bundling it together with AuthN is just asking for trouble.

Consider AWS and IAM permissions. How would you implement an IAM AuthZ policy system with OAuth? Would you actually want to?

AuthN is a relatively simple process which is why that flow through OAuth (and specifically OIDC) is fairly well trodden and defined. The OAuth Authz capabilities are infrequently used which is why you see them being so spotty.

huh? OAuth is literally a "delegated authorization framework"

It isn't. OAuth2 really did away with AuthZ and focused on AuthN, regardless of what anyone says.

You have scopes but even those outside of the OIDC scopes are wishy washy and meaningless outside of each implementation.

And this is why there's five "log in with X" buttons on every webapp, but no "tell us your IdP and we'll send you there"; let alone any kind of magical "we'll take your email address and check the DNS record of the domain to figure out who your IdP is, and then send you there."

Not because the "figuring out who to OAuth with from DNS" part would be hard to design; but because there's no way to "autodiscover" all those implementation-dependent details of how you're supposed to talk to an OAuth IdP.

Compare/contrast: SAML. SAML binding a webapp to an arbitrary IdPs just by punching in the IdP's SAML config endpoint URL? Works perfectly.

OIDC Discovery to the rescue!

https://swagger.io/docs/specification/authentication/openid-...

As much as I love OIDC, it still doesn't handle out-of-band client registration

https://datatracker.ietf.org/doc/rfc7591/

The only reason an authorisation server would support dynamic client registration, however, is because it's meant to be pluggable as the back-end for an API integration suite like Kong Enterprise.

The solutions that are one big well-known IdP have no reason to want it, so Google Auth, Amazon Cognito, and similar all don't support it.

The solutions that are aimed at letting organisations be an IdP do have reason to offer it: it makes being an IdP easier, because you can stand up a client registration service without using vendor-specific interfaces.

No-one who operates an IdP wants arbitrary, uncontrolled public client registration. It'll get abused directly, and it'll enable further abuse of the systems the IdP is meant to protect.

There's a single URL for OIDC IdPs too. It's not supported in every web app for the same reason SAML isn't: it's a two-way exchange to set up, not a one-way exchange. In OIDC (OAuth 2) you need to register a client first. In SAML you need to register a Service Provider first. Only after you've done those steps will you be able to plug the magic configuration in and have it work.

SAML is too limited to compete with OIDC/OAuth2 on the web, but it's got some advantages for enterprise authentication, so it won't be going anywhere.

Yes because embedded XMLDsig is great! I've never seen a SAML implementation that worked reliably across IDPs outside of Java, honestly.

This is my experience. It's hard because it's very complex. As a non-domain-expert it's hard to separate necessary complexity from unnecessary complexity, but it feels like there's a lot of the latter, plus enough gotchas in the former ("oh, you didn't specify an encryption algorithm that we support in your request, and our logs are terribly unhelpful for realizing that") to make it a giant pain.

Is Webauthn a better standard?

Not exactly. With webauthn the complexity lies in frontend - you have to explain how and why users should use this and what it is, how to manage enrolments and so on. Touch ID, Windows Hello and Passkeys might simplify this but it is still more complex to explain to less technical customers than, say, Face ID on an iPhone app to remember your login.

If you want an easy all-purpose login flow, sign in by clicking a link in email is still an easy fallback.

OAuth from multiple providers also has complexity - Auth0 makes it so you have to maintain separate databases for passwords if you want to support login via OAuth and login from password. You have to link accounts to login from multiple providers.

Logins are simply hard work no matter how you slice it. Eventually they will be easy but… I’m not holding my breath. :)

If you were generally comparing standards: yes. Documentation with it is a breeze.

If you're talking about a "replace one standard with the next" type of comparison, then no. You can read more about why here https://oauth.net/webauthn/

> Kerberos as a much saner SSO system than OAuth.

+1, it really solves nearly all the authn and federation problems. You still find KDC installations in places with large *nix footprints. Sometimes it's AD, sometimes it's MIT with a cross-realm trust.

It's incredibly flexible and transparent to the user. It's easy for sysadmins, and various service owners to implement as it's basically drop a keytab in place, and set an environment variable for many daemons and libraries.

IMO, the only reason it fell out of favor with the web crowd is there wasn't a gaggle of centralized providers that let them stand up services without thinking about the infrastructure. It wasn't packaged up nicely.

Kerberos just doesn't solve the same problem: it starts by giving your username and password to a fully trusted client.

OAuth was invented to avoid the need to do that. If I use some finance package and I'd like it to download my transactions from my bank, I'd really rather not have to give it my username and password, with the full access that grants including the ability to lock me out of my bank account and transfer my money. OAuth lets you instead grant permissions for that program to do something more limited, like just read your transaction history.

Kerberos was never popular with the web crowd because it's not capable of solving the problem of authorizing web applications. Before OAuth everywhere just asked for your username and password to log in on your behalf.

And the only reason Kerberos works well today is that it's always now a single-vendor solution, homogeneous across the deployment. It's an interoperability nightmare otherwise.

> Kerberos just doesn't solve the same problem: it starts by giving your username and password to a fully trusted client.

Not really. In practice, Kerberos means just loading the gssapi or sspi library, essentially using the user's login to already have the ticket-granting-ticket generated and usable for the subsequent tickets. As the client application, I never see the user's password; it's all handled for me by the OS.

The OS is the fully trusted client - but you can't repeat that trick for a web app running on a remote server.

> Kerberos just doesn't solve the same problem: it starts by giving your username and password to a fully trusted client.

That's incorrect. The client doesn't need to be trusted at all. There is also there anonymous auth, certs via pkinit (which give you yubiki and piv/cac cards as well), and otp.

> Kerberos was never popular with the web crowd because it's not capable of solving the problem of authorizing web applications.

In MS kerberos, group membership comes over with a TGT in the from of a PAC.

> Before OAuth everywhere just asked for your username and password to log in on your behalf.

I'm taking about kerberos, you'd use a TGT to log into that service.

> And the only reason Kerberos works well today is that it's always now a single-vendor solution, homogeneous across the deployment. It's an interoperability nightmare otherwise.

There are a lot of cross realm trusts between MIT/AD and Heimdall/AD, and they just work. It really isn't difficult. As I said in my original comment, it's just that it wasn't packaged up nicely. All the parts are there.

As a developer that needs to use Kerberos for authentication between hosted services, I can say that Kerberos is simply awful to debug. When you have a failed login (refuse to auth), it is so hard for the _average_ developer to debug. We have pages and pages of cookbook solutions to various problems. Some issues are purely due to Kerberos complexity, and other issues are due to each programming language and their implementation of Kerberos. Most developers have no idea how SPNEGO works (Kerberos auth over HTTPS). I spend a lot of time on these issues, and I only have a surface understanding of SPNEGO. Oh yeah, and there are many subtle issues of Microsoft Active Directory (AD) vs MIT Kerberos.

> In my experience with OAuth, one of the principle issues is that it's less a protocol and more a skeleton of a protocol.

Because it is, really.

OAuth (2.0) is really the backbone for OpenID.

OpenID Connect, specifically, where a lot of people mix those up and start looking at using OAuth for authentication. And if I'm not mistaken, logging out is implementation-dependent in OpenID Connect. There's some pseudo-standards but implementations vary in my experience. And OpenID Connect also has multiple flows for different kinds of applications. There's still a lot of confusion and complexity here.

I'm still salty about how keycloak changed their openid logout behavior, removed the old behavior in the new version when there are still a lot of oidc clients out there that still expect the old behavior. I have two inatances of keycloak using the same version and somehow both have different iodc logout behaviors. I think it's due to one instance was upgraded from older version and inherited the old behavior, but I can't get the other instance to use the old behavior (the flag mentioned in the docs didn't work) unless I downgrade the version first.

Can you elaborate on this a bit? My logout process through keycloak is through a hidden (back channel) url. As long as I hit that client url it will end the session. Applications, I find, have different behaviors. Gitea logs out the session, portainer just clears browser cookies but the session remains active

I’m using the same keycloak setup for almost 2 years now, with upgrades

I was referring to this one: https://keycloak.discourse.group/t/keycloak-redirect-uri-log...

There is a flag to restore the old behavior but it doesn't work in newer version. Strangely, an older instance of keycloak I run still uses the old behavior even after being upgraded to latest version, so this issue seems to only affect new instance only.

Yes, single log out is an ongoing nightmare. <Stares at Ping> As many here have said the size and range of use cases that OAuth and OIDC support is off its head. And that's with the big boys who have millions of users, throw in ${EveryCorp} that implements its own token server and bespoke implementation of the auth, well.. good luck to the AI trying to take over our jobs.

Dynamic client registration is standardised (https://www.rfc-editor.org/rfc/rfc7591.html), the authorization code flow is standardised (https://www.rfc-editor.org/rfc/rfc6749.html). The device authorization flow is also standardised (https://www.rfc-editor.org/rfc/rfc8628.html). The preferred desktop app authorization process is described in a best current practice document (https://www.rfc-editor.org/rfc/rfc8252.html). A metadata URL for auto-configuration of OAuth2 Authorization Servers is specified (https://www.rfc-editor.org/rfc/rfc8414.html). How to validate an access token, which also lets you know if you need to refresh it, is specified (https://www.rfc-editor.org/rfc/rfc7662.html via an API endpoint, https://www.rfc-editor.org/rfc/rfc9068.html for JWT tokens) - or just use it, and if it fails, refresh it.

It's fully specified. It's well implemented - you'd have to go out of your way to find an authorization server that doesn't do everything above, with the exception of dynamic client registration, because that's not intended for clients but rather for integration with developer portals and similar. Google Auth and Amazon Cognito don't support dynamic registration for third parties, eg, because if you're doing dynamic registration it'll be because you're operating your own AS - Okta, Auth0, and Keycloak all support it.

There's also plenty of good generic OAuth client libraries. Spring Security, the oauth2 crate for Rust, etc.

Just because the standards exist doesn't mean everybody follows them. Most and maybe all implementations have some crazy customisation. It makes the standards and docs almost worthless.

But that's hardly OAuth's fault. I can blame HTTP/2 cleartext because none of the websites I'm trying to access are working in my browser, but it's not reasonable to blame a generic protocol for not providing my niche when it's the implementations that are broken. Imagine complaining to the WHATWG because Microsoft decided to host an IRC server on port 80 and none of my browsers work.

In my experience, OAuth works great. Standard libraries Just Work, login Just Works, all you need is a URL (usually standardised) and maybe the configuration the service is expecting (i.e. the names of the permissions you're asking the user for). Usually, that's nothing special to set up. In some cases, particularly with huge "fuck you we're big enough to be the standard" vendors, there are some stupid hardcoded values and workarounds you need to deal with because they couldn't be bothered to use a library or document their flawed implementation.

I've set up a Keycloak server and authentication against it is super easy to set up. Copy-paste a domain, two generated tokens, and maybe a URL if .well-known discovery isn't implemented by the client, and that's it. Things can be easy, companies providing "OAuth" support just choose to make your life hard.

I think the problem with the spec is that it’s way too large to fully implement when you only need some secure token exchange, so everyone does only the stuff they need.

If the spec was smaller people would go all the way to be fully compliant since the marginal effort would be minimal.

For a desktop application could use the password flow[0], no need to involve a web browser. Yeah, sure, you're now handing keys to the kingdom to that application, but if it's native software it could also have installed a keylogger or used an embedded browser component and extracted the data from there.

[0] https://www.rfc-editor.org/rfc/rfc6749#section-4.3

OAuth isn't an SSO system :P

Reply on news.ycombinator.com