Re: [GNAP] Consensus Call on Continuation Request

[Stephen Moore <[e-mail]>]

Hi Fabien,

For #3) Even after I typed out the hypothetical attack, that was sort of in
the back of my mind, it isn't a huge risk there. So I actually agree with
Dick there. Something doesn't sit right with me for the unique URL
solution, so I don't like it and came up with a hypothetical that seems
like it could be a down side.

I still think the access token model with the signed request is the way I'd
like to go, because again, it's a mechanism I'd be implementing anyway to
talk to any 'normal' resource. The fact is there is _something_
representing context that has to pass back and forth here, whether that is
an access token (which I feel like is more flexible for extensions etc), a
unique url, or even a cookie sent in the cookie header. So just to
re-iterate, I'm a +1 on this pull request, speaking as a lazy developer ;)
-steve

On Fri, Dec 11, 2020 at 8:51 PM Fabien Imbault <[e-mail]>
wrote:

> Again speaking in my own name here.
>
> Dick, we know you'd prefer to have a different design, but this PR
> shouldn't be about that.
>
> Back on your 3 items :
>
> 1) yes we could make pre-register mandatory, but we already decided that
> wouldn't be how that would work. We have a client instance that allows a
> more generic and flexible pattern (which BTW also allows what you want)
>
> 2) instead of blame arguments of who's less restful/HATEOAS/whatever that
> have the tenancy to flame conversations, I suggest we speak in less
> abstract terms and ask ourselves what that means in practice for devs.
> Stephen and several others (myself included) have expressed that it
> wouldn't be harder to implement, it would even simplify things quite a lot.
> If you disagree please send us a code sample to really show that point by
> example, because that's really not obvious.
>
> 3) "If someone has the client credentials, they can impersonate the
> client, and all bets are off." Are you seriously making this argument?
> Because if you have a better proposal than using cryptographic keys, I'm
> all hears. You make it look like there's a problem, while in reality we're
> only relying on the basic assumption of all modern digital communications.
>
> And more importantly you never responded to the issues of how to avoid the
> security pitfalls of what you proposed.
>
> Fabien
>
>
> Le sam. 12 déc. 2020 à 00:35, Dick Hardt <[e-mail]> a écrit :
>
>>
>>
>> On Fri, Dec 11, 2020 at 2:53 PM Stephen Moore <[e-mail]> wrote:
>>
>>> But from the spec:
>>> "
>>> When sending a non-continuation request to the AS, the RC MUST identify
>>> itself by including the client field of the request...
>>> ...
>>> key (object / string) : The public key of the RC to be used in this
>>> request as described in {{request-key}}. This field is REQUIRED.
>>> ...
>>> "
>>>
>> So on the initial request, the key will be there.
>>>
>>
>> The client field can be an object or a string. If the client is
>> pre-registered, then a string could be provided instead of an object.
>>
>>>
>>
>>>
>>> If you don't have the access token, then how do you differentiate
>>> between two requests from the same web application by two different users?
>>> Is the web application supposed to have different credentials for every
>>> request?
>>>
>>
>> The AS returns a URI for manipulating the request. I would change the
>> spec so that each request would have a unique URI. This is the usually
>> RESTful pattern that the resource (the grant request) has an URI.
>>
>>
>>
>>> So in this case, the easy way out is to pass the access token to the
>>> client, who then, as i stated before, treats the continue request as a RS
>>> call (albeit a specialized version of the RS where the RS is the AS) OR to
>>> use the unique URL,
>>> but that seems open to a brute force attack by a malicious RC. (What
>>> would be the point of that attack, I don't know, I guess if someone had the
>>> client credentials but not any subjects/resources they could try to
>>> intercept the grant via continue... I just don't feel right locking things
>>> down to unique URLs that way.)
>>>
>>
>> If someone has the client credentials, they can impersonate the client,
>> and all bets are off.
>>
>> LOTS of RS servers return a resource specific URL -- my proposal is no
>> different.
>>
>>
>>
>>
>>> -steve
>>>
>>> On Fri, Dec 11, 2020 at 5:33 PM Dick Hardt <[e-mail]> wrote:
>>>
>>>> Hi Stephen
>>>>
>>>> The client is signing the first request. The key *might* be in the
>>>> body. The client is signing all the subsequent requests as well. The
>>>> "access token" is not needed by the client to prove it is authorized as the
>>>> client is proving it is the same client again.
>>>>
>>>> In other words, I don't see the need for an access token, so it does
>>>> not need to be put in a URL or an auth header.
>>>>
>>>> If a developer really, really wants to hand context back to the client
>>>> for subsequent calls, they can put it in the URL or some other method.
>>>> Putting it in the HTTP Authorization header is confusing because it is NOT
>>>> an access token -- it is the context of the request.
>>>>
>>>> ᐧ
>>>>
>>>> On Fri, Dec 11, 2020 at 2:01 PM Stephen Moore <[e-mail]>
>>>> wrote:
>>>>
>>>>> Even though I've only been lightly following things, I feel the need
>>>>> to voice my preference as a developer since I will probably someday have to
>>>>> either write a RC or RS...
>>>>>
>>>>> The way I see it is the RC makes the initial request to the AS as part
>>>>> of this request, it provides it's key in the body... (So no use of the
>>>>> Authorization header)
>>>>> At this point that request, represented by the continue URL + "Access
>>>>> Token", from my lazy developer standpoint, is a Resource Endpoint and
>>>>> Access Token, and the AS is acting as a specialized RS in this case.
>>>>> So my client posts to whatever URL with the 'access token' in the
>>>>> authorization header, just like acting on any other resource I have a token
>>>>> for. YES, I get a new token value to use every call, and there is a
>>>>> decision point of "Do I have another continue, or do I have a real token
>>>>> for the resource..." But the mechanism is the same to me in the client.
>>>>> Personally I like that, because if I have an access_token, I already
>>>>> think "Put it in the auth header."
>>>>>
>>>>> So my vote would be +1 for the pull request at this time.
>>>>> -steve
>>>>>
>>>>> On Fri, Dec 11, 2020 at 3:04 PM Dick Hardt <[e-mail]>
>>>>> wrote:
>>>>>
>>>>>> inline ...
>>>>>>
>>>>>> On Fri, Dec 11, 2020 at 9:58 AM Justin Richer <[e-mail]>
>>>>>> wrote:
>>>>>>
>>>>>>> Others had already responded to this previous thread, but I wanted
>>>>>>> to add a couple points to clarify some things.
>>>>>>>
>>>>>>> 3) What the client has to do with the "access token" is not the same
>>>>>>> as access tokens for an RS. The client gets a new "access token" for each
>>>>>>> grant request, and for each API call to the AS, and the client learns it
>>>>>>> can not make any more API calls for that specific request when it does not
>>>>>>> get an "access token" back. This is a completely different design pattern
>>>>>>> than calling an RS API with an access token, and is a new design pattern
>>>>>>> for calling APIs. This adds complexity to the client that it would not
>>>>>>> normally have, and I don't think GNAP is the right place to start a new
>>>>>>> design pattern.
>>>>>>>
>>>>>>>
>>>>>>> I’m not sure what you mean by these being different — the whole
>>>>>>> point of the design is that the client would be doing the same thing with
>>>>>>> the access token at the AS that it does with the RS by re-using the access
>>>>>>> token structure. Can you please describe what the differences are, apart
>>>>>>> from the rotation? Presentation of the token and signing of the message are
>>>>>>> identical.
>>>>>>>
>>>>>>
>>>>>> The client is getting the "access token" from its API. It is not
>>>>>> using an "access_token" in other API calls to the AS.
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> Rotation of the access token and artifacts for ongoing continuation
>>>>>>> responses is a separate issue to be discussed:
>>>>>>> https://github.com/ietf-wg-gnap/gnap-core-protocol/issues/87
>>>>>>>
>>>>>>> And for what it’s worth, GNAP is absolutely the right place to have
>>>>>>> new designs — not that this is one.
>>>>>>>
>>>>>>
>>>>>> You are proposing a new way for an API to provide context for
>>>>>> subsequent API calls. Looks out of scope to me.
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> 4) Clients that only want claims from the AS and no access tokens
>>>>>>> will be required to support an API calling mechanism they would not have to
>>>>>>> support otherwise.
>>>>>>>
>>>>>>>
>>>>>>> Correct, but the delta between the calls a client would make with
>>>>>>> and without an access token is vanishingly small. The client has to sign
>>>>>>> the initial request in some fashion, and it will sign the continuation
>>>>>>> request in the same exact fashion, but now include an access token in that
>>>>>>> request.
>>>>>>>
>>>>>>
>>>>>> Per my other point, there is no value to me in my implementations of
>>>>>> passing context back and forth between the client and AS -- so it is extra
>>>>>> work providing no value.
>>>>>>
>>>>>> Also, any client authentication mechanism that wants to use the HTTP
>>>>>> Authentication header is precluded from using it.
>>>>>>
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> Clients making a request to an AS and not getting an access token is
>>>>>>> a new design pattern. I think it has value and should be included, but
>>>>>>> OAuth today shows us the immense value of getting access tokens for calling
>>>>>>> APIs, and so we shouldn’t optimize away from that pattern.
>>>>>>>
>>>>>>>
>>>>>>> 5) If the AS does not provide an "access token", there is no
>>>>>>> mechanism for a client to delete the request, as the client is not allowed
>>>>>>> to make a call without an "access token".
>>>>>>>
>>>>>>>
>>>>>>> More properly, if the AS does not provide a “continue” field then
>>>>>>> the client can’t delete the request — and yes, that’s intentional. The AS
>>>>>>> is telling this client instance that it can’t do anything else with this
>>>>>>> ongoing request. If the AS wants to allow the client to manage it, it will
>>>>>>> include the mechanisms to do so in the “continue” field.
>>>>>>>
>>>>>>
>>>>>> There is nuance in that intention. A related concern is that deleting
>>>>>> a request does not seem like it is a "continue" operation.
>>>>>>
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 6) There is no standard identifier for the request. Debugging and
>>>>>>> auditing are hampered by the client and AS having no standard way to
>>>>>>> identifying a request. While one AS may provide a unique URL for each grant
>>>>>>> request, another AS may use a persistent "access token" to identify the
>>>>>>> grant request, and other ASs may issue a new "access token" on each API
>>>>>>> call, providing no persistent identifier for the request.
>>>>>>>
>>>>>>>
>>>>>>> Debugging and auditing this kind of thing are functions of the AS.
>>>>>>> How is interoperability harmed by different ASs having different methods to
>>>>>>> identify their internal data elements? The client doesn’t need any
>>>>>>> knowledge of the AS’s identifiers, it just needs to know the next steps for
>>>>>>> continuing the negotiation.
>>>>>>>
>>>>>>
>>>>>> Debugging between the client and the AS was what I was referring to.
>>>>>> How does a client developer identify the request when communicating to the
>>>>>> AS developer. Seems complicated.
>>>>>>
>>>>>> ᐧ
>>>>>> --
>>>>>> TXAuth mailing list
>>>>>> [e-mail]
>>>>>> https://www.ietf.org/mailman/listinfo/txauth
>>>>>>
>>>>> ᐧ
>> --
>> TXAuth mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/txauth
>>
>