Re: [OAUTH-WG] Questions on draft-ietf-oauth-dyn-reg-09 - token_endpoint_auth_method

[Justin Richer <[e-mail]>]

If you want your client app to be able to speak to all servers, yes, 
you'll have to implement all methods. But even this is a problem because 
servers can come up with a new method -- that's the whole point of 
making this field extensible, isn't it? And again, you could make the 
same argument about grant_type and several other fields. Say you've got 
a client that only does authorization_code, but the server sends back 
"implicit" to the client. Will they be able to speak? No, and if the 
client application or library wants to be able to speak to this server 
it's going to have to implement the implicit flow to do so.

So I think that your concern here, while definitely valid, is a bit 
misleading. I don't think it's OAuth Dyn Reg's goal to solve the 
complete cold-boot problem across all possible combinations of clients 
and servers, since a client application is going to need to be able to 
speak whatever protocol OAuth is protecting anyway in order to do 
anything past the authorization step. However, I do think it makes 
perfect sense for a client library to be able to say "here's the value 
that I want" and have the server say "here's the value that you get" in 
a means that's parallel across all of these fields.

  -- Justin

On 04/25/2013 01:09 PM, Phil Hunt wrote:
> If the server is free to ignore, doesn't that mean the client ends up 
> having to implement *ALL* methods anyway? In the case of Connect, the 
> client has to implement all methods supported by Connect because it 
> *could* be forced by any particular deployer to use a particular method.
>
> The dynamic draft + unspecified discovery doesn't (as described) 
> doesn't make clients or servers actually easier to implement.  Any 
> client not implementing all methods runs the risk of sooner or later 
> not being compatible.
>
> I agree there is a need for registration to convey to the client what 
> authentication method and client credential it has been issued.
>
> Phil
>
> @independentid
> www.independentid.com <http://www.independentid.com>
> [e-mail] <mailto:[e-mail]>
>
>
>
>
>
> On 2013-04-25, at 9:35 AM, Justin Richer wrote:
>
>>
>> On 04/25/2013 12:23 PM, Phil Hunt wrote:
>>> I am not sure what to think about this.
>>>
>>> 1. I don't want this spec to be incomplete because there is another 
>>> external spec published for a specific purpose.
>> If we needed a registry we could define it in this spec. If we chose 
>> a non-registry method (like how grant_type uses URIs) then it's 
>> self-contained. I prefer the latter, myself.
>>
>>> 2. In general i would like to avoid negotiations. It seems to me 
>>> that any developer making a client for a mutli-site deployed service 
>>> would already be aware of what the generic service require and 
>>> implement based on that oob knowledge. Thus when the server merely 
>>> informs, the client should be prepared.
>> The client already has to be prepared because the server will inform, 
>> though. But by making it bidirectional (like all the other 
>> parameters) it gives the client a chance to ask based on whatever 
>> information it wants to. The server is perfectly free to ignore what 
>> the client asks for and just hand it something, just as it's free to 
>> do so for all fields.
>>
>>> IMHO, though i understand the motivation, negotiation of client auth 
>>> creates more problems then it seems to solve in this case. Downgrade 
>>> attack is just one of the problems.
>> The parameters of the negotiation are going to be 
>> application-protocol specific and depend on things like a Discovery 
>> layer for them to be fully dynamic from a cold boot. And even then 
>> servers and clients that care about this value are going to have to 
>> know how to enforce it (it's the same as it is with grant_type, 
>> response_type, scope, and others).
>>
>> If nothing else, it lets servers tell clients which auth method to 
>> use at the token endpoint. The client can be smart and try to figure 
>> out if it actually supports that method before trying it, or it can 
>> be dumb and ignore the returned value. But with this parameter in the 
>> protocol, both sides at least have a *chance* of doing the right thing.
>>
>>  -- Justin
>>
>>>
>>> Phil
>>>
>>> On 2013-04-25, at 7:21, John Bradley <[e-mail] 
>>> <mailto:[e-mail]>> wrote:
>>>
>>>> I am OK with having a registry,  Connect profiles the jwt 
>>>> assertions draft to add specifics around keying material for 
>>>> interoperability.
>>>> It may well be that people will want other methods including SAML 
>>>> assertions.
>>>>
>>>> While sending a list of methods the client supports to the server 
>>>> and getting back what is supported by the server works, it is awkward.
>>>>
>>>> My point is mostly that there are lots of other parameters that are 
>>>> set only by the server.  I don't know if we want to start adding 
>>>> them to registration just to allow for them to be discovered.   
>>>> Using a simple .well-known discovery document seems simpler.
>>>>
>>>> So making token_endpoint_auth_method an array is not the end of the 
>>>> world and would work, but heads us in what I think is the wrong 
>>>> direction.
>>>>
>>>> John B.
>>>>
>>>> On 2013-04-25, at 10:41 AM, Justin Richer <[e-mail] 
>>>> <mailto:[e-mail]>> wrote:
>>>>
>>>>> John, I don't think that anybody's actually suggesting that we add 
>>>>> server discovery in here. :) But since the server does echo back 
>>>>> the configuration in its response, the client is discovering a few 
>>>>> things at run time about what it can and can't do with a 
>>>>> particular server. It's entirely possible that the client might 
>>>>> get back a configuration option that it can't use (say, it can't 
>>>>> do client_secret_jwt assertions but it can do
>>>>>
>>>>> From what I can see from this thread though there are two open 
>>>>> questions that Phil's raised:
>>>>>
>>>>>
>>>>>  1: Extensibility of token_endpoint_auth_method values, and where 
>>>>> potential values are listed (IANA? fully qualified URIs for things 
>>>>> not in the base spec?)
>>>>>
>>>>>  2: Plurality of token_endpoint_auth_method (could be left as a 
>>>>> string, made an array, or be a JSON-style optional plural: string 
>>>>> value if singular or array if plural)
>>>>>
>>>>>
>>>>>
>>>>> I think the first one should be addressed, but we just need to 
>>>>> pick the method. I'm personally in favor of the same method we 
>>>>> used for grant_type, which is short values in the spec and 
>>>>> extensions as fully qualified URIs. The second one could break 
>>>>> existing implementations (and other dependent specs), so if we 
>>>>> change it, it has to be for very good reason.
>>>>>
>>>>>  -- Justin
>>>>>
>>>>> On 04/24/2013 07:23 PM, John Bradley wrote:
>>>>>> In Connect there is a AS discovery before registration.
>>>>>>
>>>>>> The general pattern is the RP discovers the capabilities of the 
>>>>>> AS authentication methods and algorithms supported by the AS.
>>>>>> The client then picks the best options for it and registers them.
>>>>>>
>>>>>> It would in theory work of the client knowing nothing about the 
>>>>>> AS pushed it's capabilities at the AS as you are suggesting and 
>>>>>> let the AS pick.
>>>>>>
>>>>>> My general feeling is that discovery with the client picking the 
>>>>>> options works best.
>>>>>>
>>>>>> In many cases the client doesn't need to register parameters as 
>>>>>> they can be selected at run time once it knows what a server 
>>>>>> supports.
>>>>>> The token endpoint authentication method was a bit of a special 
>>>>>> case where even though it could be all dynamic and work, you do 
>>>>>> want to register a choice to prevent backwards compatibility attacks.
>>>>>>
>>>>>> I don't really want to complicate registration by trying to make 
>>>>>> it also cover AS discovery.
>>>>>>
>>>>>> John B.
>>>>>>
>>>>>> On 2013-04-24, at 7:55 PM, Phil Hunt <[e-mail] 
>>>>>> <mailto:[e-mail]>> wrote:
>>>>>>
>>>>>>> Right and if the client wants a method not supported then what?
>>>>>>>
>>>>>>> Why can't the client offer up a list of methods it is able to 
>>>>>>> support, say in order of preference?
>>>>>>>
>>>>>>> The text appears to indicate only one value may be passed.
>>>>>>>
>>>>>>> Given the way it is written. It seems better to just have the 
>>>>>>> server say the client must do authn method X in the response.
>>>>>>>
>>>>>>> Phil
>>>>>>>
>>>>>>> @independentid
>>>>>>> www.independentid.com <http://www.independentid.com/>
>>>>>>> [e-mail] <mailto:[e-mail]>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On 2013-04-24, at 3:41 PM, John Bradley wrote:
>>>>>>>
>>>>>>>> In Connect the AS may support a number of token endpoint 
>>>>>>>> authentication methods. The reason to allow a client to 
>>>>>>>> register using a particular one is to prevent downgrade attacks.
>>>>>>>>
>>>>>>>> If the client wants to always use an asymmetric signature you 
>>>>>>>> don't want to allow attackers to use weaker methods like http 
>>>>>>>> basic.
>>>>>>>>
>>>>>>>> So a server may support any number of methods, but it is 
>>>>>>>> reasonable for a client to specify which one it is going to 
>>>>>>>> use. In a closed system that may not be that useful but in a 
>>>>>>>> open system where the AS has a looser relationship to the 
>>>>>>>> client it is important.
>>>>>>>>
>>>>>>>> John B.
>>>>>>>>
>>>>>>>> On 2013-04-24, at 7:30 PM, Phil Hunt <[e-mail] 
>>>>>>>> <mailto:[e-mail]>> wrote:
>>>>>>>>
>>>>>>>>> Hmmm… what was the objective or use case for having the client 
>>>>>>>>> being able to choose in the first place?
>>>>>>>>>
>>>>>>>>> It seems to me that the AS will make a decision based on many 
>>>>>>>>> factors. As you say, there isn't any other place that 
>>>>>>>>> enumerates the various [authn] methods a client can use to 
>>>>>>>>> access the token endpoint.  So, why do it?
>>>>>>>>>
>>>>>>>>> Phil
>>>>>>>>>
>>>>>>>>> @independentid
>>>>>>>>> www.independentid.com <http://www.independentid.com/>
>>>>>>>>> [e-mail] <mailto:[e-mail]>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 2013-04-24, at 2:07 PM, Justin Richer wrote:
>>>>>>>>>
>>>>>>>>>> Seems reasonable to me, can you suggest language to add in 
>>>>>>>>>> the capability? Would it require an IANA registry? Right now 
>>>>>>>>>> there isn't any other place that enumerates the various 
>>>>>>>>>> methods that a client can use to access the token endpoint.
>>>>>>>>>>
>>>>>>>>>>  -- Justin
>>>>>>>>>>
>>>>>>>>>> On 04/24/2013 04:17 PM, Phil Hunt wrote:
>>>>>>>>>>> For parameters to token_endpoint_auth_method, the spec has 
>>>>>>>>>>> defined "client_secret_jwt" and "private_key_jwt". Shouldn't 
>>>>>>>>>>> there be similar options of SAML?
>>>>>>>>>>>
>>>>>>>>>>> Shouldn't there be an extension point for other methods?
>>>>>>>>>>>
>>>>>>>>>>> Phil
>>>>>>>>>>>
>>>>>>>>>>> @independentid
>>>>>>>>>>> www.independentid.com <http://www.independentid.com/>
>>>>>>>>>>> [e-mail] <mailto:[e-mail]>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> _______________________________________________
>>>>>>>>>>> OAuth mailing list
>>>>>>>>>>> [e-mail]
>>>>>>>>>>> https://www.ietf.org/mailman/listinfo/oauth
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>>>>>> OAuth mailing list
>>>>>>>>> [e-mail] <mailto:[e-mail]>
>>>>>>>>> https://www.ietf.org/mailman/listinfo/oauth
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>
>