Re: [Txauth] Use Case: Directed Tokens

Justin Richer <jricher@mit.edu> Tue, 07 July 2020 20:39 UTC

Return-Path: <jricher@mit.edu>
X-Original-To: txauth@ietfa.amsl.com
Delivered-To: txauth@ietfa.amsl.com
Received: from localhost (localhost [127.0.0.1]) by ietfa.amsl.com (Postfix) with ESMTP id F2B183A0A62 for <txauth@ietfa.amsl.com>; Tue, 7 Jul 2020 13:39:34 -0700 (PDT)
X-Virus-Scanned: amavisd-new at amsl.com
X-Spam-Flag: NO
X-Spam-Score: -1.886
X-Spam-Level:
X-Spam-Status: No, score=-1.886 tagged_above=-999 required=5 tests=[BAYES_00=-1.9, HTML_MESSAGE=0.001, RCVD_IN_MSPIKE_H4=0.001, RCVD_IN_MSPIKE_WL=0.001, SPF_HELO_NONE=0.001, SPF_PASS=-0.001, T_REMOTE_IMAGE=0.01, URIBL_BLOCKED=0.001] autolearn=ham autolearn_force=no
Received: from mail.ietf.org ([4.31.198.44]) by localhost (ietfa.amsl.com [127.0.0.1]) (amavisd-new, port 10024) with ESMTP id w7cZwvsiPrxF for <txauth@ietfa.amsl.com>; Tue, 7 Jul 2020 13:39:31 -0700 (PDT)
Received: from outgoing.mit.edu (outgoing-auth-1.mit.edu [18.9.28.11]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ietfa.amsl.com (Postfix) with ESMTPS id A75843A0A56 for <txauth@ietf.org>; Tue, 7 Jul 2020 13:39:30 -0700 (PDT)
Received: from [192.168.1.7] (static-71-174-62-56.bstnma.fios.verizon.net [71.174.62.56]) (authenticated bits=0) (User authenticated as jricher@ATHENA.MIT.EDU) by outgoing.mit.edu (8.14.7/8.12.4) with ESMTP id 067KdSTS022749 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Tue, 7 Jul 2020 16:39:29 -0400
From: Justin Richer <jricher@mit.edu>
Message-Id: <6846D2F0-5096-404B-A529-676146392F45@mit.edu>
Content-Type: multipart/alternative; boundary="Apple-Mail=_2EC0B3C3-DF0A-4896-ADD2-3A8885D9C3D1"
Mime-Version: 1.0 (Mac OS X Mail 13.4 \(3608.80.23.2.2\))
Date: Tue, 07 Jul 2020 16:39:28 -0400
In-Reply-To: <6CCD515B-BE87-452B-9034-777D90E110DD@mit.edu>
Cc: Tobias Looker <tobias.looker@mattr.global>
To: txauth@ietf.org
References: <4F145676-A126-4D35-8890-A0DDF891EA06@mit.edu> <32ae1a93-fc9d-cd15-798e-ec493482dd26@free.fr> <90F181EE-8E34-4486-BCFB-ADACE55A55CF@mit.edu> <dd8ef917-c63a-0070-810a-aecfd9aac0a0@free.fr> <CAJmmfSRMWRMQbfZ2ktaRRq1oVeZtXSRf0TGiCJmcLi1FJF6N+w@mail.gmail.com> <6CCD515B-BE87-452B-9034-777D90E110DD@mit.edu>
X-Mailer: Apple Mail (2.3608.80.23.2.2)
Archived-At: <https://mailarchive.ietf.org/arch/msg/txauth/Y_cVLYN3zEPKcqcSA9OGtVh-llQ>
Subject: Re: [Txauth] Use Case: Directed Tokens
X-BeenThere: txauth@ietf.org
X-Mailman-Version: 2.1.29
Precedence: list
List-Id: <txauth.ietf.org>
List-Unsubscribe: <https://www.ietf.org/mailman/options/txauth>, <mailto:txauth-request@ietf.org?subject=unsubscribe>
List-Archive: <https://mailarchive.ietf.org/arch/browse/txauth/>
List-Post: <mailto:txauth@ietf.org>
List-Help: <mailto:txauth-request@ietf.org?subject=help>
List-Subscribe: <https://www.ietf.org/mailman/listinfo/txauth>, <mailto:txauth-request@ietf.org?subject=subscribe>
X-List-Received-Date: Tue, 07 Jul 2020 20:39:35 -0000

I’ve been thinking more about these use cases, and it might help us to think of GNAP’s access tokens as capabilities in the computer security sense — but more specifically, capabilities that are delivered with their own context for use. In that way, an otherwise naive client can be handed an access token and simply know exactly what to do with it. This context would be delivered alongside the token, so the token material itself remains opaque to the client.

I wanted to bring up a W3C spec for a capabilities model based on linked data:

https://w3c-ccg.github.io/zcap-ld/ <https://w3c-ccg.github.io/zcap-ld/>

Building a fully featured capabilities context is difficult, at the very least. And unfortunately, I don’t think this spec actually gives us any viable solutions to work with. In particular the “actions” section is effectively blank, offloading the work to an external JSONLD process (side note, this is one of the problems I have with specs based on JSONLD, they externalize the important parts into local contexts and break interoperability — but I digress). But at least it’s another way of looking at the problem space that might be outside the familiar zone of many of the OAuth world.

 — Justin

> On Jun 26, 2020, at 9:23 AM, Justin Richer <jricher@mit.edu> wrote:
> 
> On Jun 25, 2020, at 9:07 PM, Tobias Looker <tobias.looker@mattr.global <mailto:tobias.looker@mattr.global>> wrote:
>> 
>> I find this feature interesting and think it could be an important pattern going forward to enable an AS to be able to describe the utility of a token to the client, however as already pointed out in the thread I think there is some potential hidden complexity that would need to be ironed out such that it does not make the simple things complicated.
>> 
>> Justin, do you see this feature as something the client has to explicitly request, i.e "please give me access and instructions on how to use my access" or rather that the AS would just return this information in conjunction with the access token if it had it available?
>> 
> 
> This is something that I’d brought up as a possibility on the previous thread — something like a flag the client would set. This would be especially important if the AS wants to return multiple access tokens but the client requested 1, or the client requested N and the AS wants to return M in response. Basically any time there’s a mismatch, that’s extra complexity on the client that I really don’t think we want to be universal. A flag to turn that kind of direction and splitting on would be a potential start. But there are potential snags here too, and it comes down to how you manage the defaults...
> 
>> > This is just the start, too. Things get even more complex if the client can ask for multiple different kinds of resources at once. What if the AS decides that the client needs a hyper-focused directed token for one part of the API, but can use a general token for other stuff? Can it signal that to the client? And if it can, does that mean that all clients need to be prepared for that kind of thing?
>> 
>> Would a potential default be that if a client did for any reason not support processing the additional information returned with the access_token, just to ignore it? I think in the spirit of keeping the simple things simple, this potentially makes sense?
> 
> That’s the real trick, if you ask me. It has to be :safe: for a client to ignore this information. Which means the AS can’t rely on the client “doing the right thing” with the information in the “token directions” portion of the response. Even setting aside the advanced and related “split tokens” idea above, we need to make sure that an AS/RS setup is built such that if the AS tells a client “only do X with your token” and the client does “Y”, then there are other protections and practices in place to prevent things from going haywire if the client does “Y” either by accident or through ignorance. 
> 
> If OAuth 2 has taught us anything, it’s that client applications are going to be the laziest participants in the security model. And that makes sense, really — security isn’t what the client apps are trying to do, they’re trying to use the RS to do something. So we need to make sure that whatever system we design will fail on the safe side as much as possible, and keep things simple for the client.
> 
>> 
>> > here are some worked out samples of this:
>> https://wiki.idesg.org/wiki/index.php/High_Assurance_AZ_Token <https://wiki.idesg.org/wiki/index.php/High_Assurance_AZ_Token>
>> https://mattrglobal.github.io/oidc-client-bound-assertions-spec/ <https://mattrglobal.github.io/oidc-client-bound-assertions-spec/>
>> Peace ..tom
>> 
>> As one of the authors of those mentioned drafts, I am interested in discussing that, but perhaps on a seperate thread? As at least the bound assertion spec is primarily concerned with binding mechanisms for the artifacts produced from an authorization flow (specifically identity claims), whereas I think directed access tokens is just more generally talking about whether GNAP should support an AS conveying how to use an access_token it produces during a flow with a client.
>> 
> 
> I think that these are separate issues as well.
> 
>  — Justin
> 
>> Thanks,
>>  <https://mattr.global/>	 	
>> Tobias Looker
>> Mattr
>> +64 (0) 27 378 0461
>> tobias.looker@mattr.global <mailto:tobias.looker@mattr.global>
>>  <https://mattr.global/>	 <https://www.linkedin.com/company/mattrglobal>	 <https://twitter.com/mattrglobal>	 <https://github.com/mattrglobal>
>> This communication, including any attachments, is confidential. If you are not the intended recipient, you should not read it - please contact me immediately, destroy it, and do not copy or use any part of this communication or disclose anything about it. Thank you. Please note that this communication does not designate an information system for the purposes of the Electronic Transactions Act 2002.
>> 
>> 
>> On Wed, Jun 24, 2020 at 10:14 PM Denis <denis.ietf@free.fr <mailto:denis.ietf@free.fr>> wrote:
>> Justin, I fear we are still far away from an agreement about what this BoF should do.
>> 
>> Tom Jones is saying " I am getting tired of the whack-a-mole approach ..."
>> 
>> I don't agree with you when you write: "I think it’s going to be overwhelmingly common that a given RS is going to trust exactly one AS 
>> that it knows about ahead of time". Such an architecture would have exactly the same limitations as OAuth 2.0. and would not be scalable.
>> 
>> Before we start, we should have a clear view of the whole picture rather than starting from one scenario and expanding it in many different directions.
>> For building an architecture, a pre-requirement is to define ALL the trust relationships. I don't believe that we have an agreement at this time on what 
>> these trust relationships are. 
>> 
>> You are also using the following wording: "methods for the AS and RS to communicate what the token’s good for". 
>> With such a paradigm, it would be impossible to protect the user's privacy. 
>> 
>> A key question is : Will GNAP take care of the user's privacy or will GNAP prevent to take care of the user's privacy ?
>> 
>> About "the ability for the client to get several access tokens to get different resources from one request" is indeed a complex case.
>> 
>> No one (including ASs) is able to predict in advance which access tokens will be needed, since they depend upon the kind of operation(s) 
>> the client will be willing to perform. For privacy reasons, ASs should be kept ignorant of all the operations that a client is going to perform 
>> on one or more resource servers. I believe that every effort should be made to avoid the AS to be in a position to be able to trace the operations 
>> performed by its clients on various RSs.
>> 
>> To handle the complex case you envision, the full workflow of operations needs to be considered with a detailed focus on the time 
>> at which the user's consent and choice happens (consent and choice is the first privacy principle from ISO 29100).
>> 
>> First of all, an access token could be targeted to a service rather than to a server. This would already solve many cases.
>> 
>> When a RS needs to call another RS (which does not support the same service) then the client should be informed by the first RS.
>> His "consent and choice" should then be obtained by the first RS and, when the user agrees, the client should then ask an access token 
>> to one of the ASs trusted by the second RS in order to perform the subsequent operation.  
>> 
>> Denis
>> 
>> PS.  In an email sent on June 23 you wrote: " I think an on-device AS is an important use case".  I am sorry, but I don't understand.
>>        However, I do understand what "a server-based AS" is.
>> 
>> 
>>> Denis, thanks for the great comments. I think that your trust model is pretty different from what I’d laid out initially, and while it’s an interesting and important one, it is not what I meant by “multiple access tokens” in my discussion below. I think that might be the cause of some disconnect here, but that doesn’t mean it’s out of reach for what the WG is after.
>>> 
>>> When I refer to multiple access tokens, and what the charter calls out as multiple access tokens, is the ability for the client to get several access tokens to get different resources from one request. I personally see this as an optimization of a specific set of use cases, some of which I discussed in my original email. There are others, I’m sure, but all the ones I’ve seen are around the client needing to get several different kinds of access through an AS. 
>>> 
>>> I think it’s going to be overwhelmingly common that a given RS is going to trust exactly one AS that it knows about ahead of time. But for RS’s that can trust multiple AS’s, then we should have a model that can accommodate that as well. That’s why the charter calls out methods for the AS and RS to communicate what the token’s good for. I think the basis of those methods is going to start with a common token model, and likely incorporate into things like token formats and introspection-style token APIs. 
>>> 
>>>  — Justin
>>> 
>>>> On Jun 22, 2020, at 3:55 AM, Denis <denis.ietf@free.fr <mailto:denis.ietf@free.fr>> wrote:
>>>> 
>>>> Hello Justin,
>>>> 
>>>> A few comments between the lines.
>>>> 
>>>>> One of the most important aspects to GNAP is going to be an ability to address things that OAuth 2 can’t, or at least can’t do without significant gymnastics. So with that, I wanted to bring back a use case discussion that originally came up while we were talking about the possibility of multiple access tokens, a few months back. I don’t know if this concept has a regular term, so I’m going to call it “directed access tokens” until we come up with something better — assuming we decide this is worthwhile. 
>>>> I don't understand what may mean "directed access tokens” but I understand what means "multiple access tokens". 
>>>> When speaking of "multiple access tokens", these access tokens may come from one or more ASs.
>>>> 
>>>>> In OAuth, the client’s supposed to always know where to send its access token, and that knowledge is completely outside the protocol. This makes a lot of sense for many (if not most) deployments, as OAuth is really there to enable the API access that the client already knows about.
>>>>> 
>>>>> But we’re now in a world where a client could be talking to a generic API that could be deployed at a number of different places, or a cloud deployment that the AS wants to be able to dispatch the client to. 
>>>> As soon the AS is placed (again) at the centre of the model, the AS will have the ability to act as "Big Brother".
>>>> OAuth 2.0 has not taken care of privacy issues. On the contrary, GNAP should take care of them.
>>>> 
>>>>> In order to do this, the AS needs to be able to communicate to the client not only the token information (and any related key and presentation information), but also a set of directions about what that specific token is good for. This needs to be information outside the token itself, since I believe we want to keep OAuth 2’s feature of having the token be opaque to the client. Note: while we could map all of these to different resource requests (in XYZ parlance) or scopes (in OAuth 2 legacy parlance) on the request side, this isn’t enough to tell the client what to do with the token if it doesn’t know already. 
>>>>> 
>>>>> I know of two use cases already in the wild, where people are addressing things using a mix of existing standards and some proprietary extensions to address things within their silos. I’ll try to summarize here, but I know the folks I’ve been talking to about this are also on the list so hopefully they can chime in with more detail or any corrections for something I’ve missed. 
>>>>> 
>>>>> (1) The client knows what resource it’s calling, but it doesn’t know where it’s hosted. Everything is in a single security domain controlled by the AS, 
>>>> Speaking of "multiple access tokens" is in contradiction with single security domain "controlled" by an AS. 
>>>> A user may need to demonstrate some of his identity attributes granted e.g. by his bank but may also 
>>>> need to demonstrate one or more diplomas granted by one or more universities. The bank cannot be 
>>>> the "primary issuer" of a university diploma. It should not be either a "secondary issuer", because 
>>>> the bank does not "need to know" what are the diplomas of the user.
>>>> 
>>>>> but the AS needs to decide at runtime which specific instance of the API to direct the client to. Since things are closely tied together, the client just needs to effectively known an identifier for the RS, and this is currently implemented as a URI. Once the client has that identifier, it knows how to dispatch that token to that instance of the RS.
>>>> There is no good reason why the AS should be involved in that step. 
>>>> OAuth 2.0 is/was implicitly mandating a strong relationship between ASs and RSs which makes it non scalable.
>>>> GNAP should be based on a simple trust relationship : a given RS trusts some ASs for access tokens that contains some types of attributes.
>>>> An AS should not need to know in advance (or even at the time of an access token request) the RSs that are trusting it.
>>>> 
>>>> A client could first talk to a "global service provider" which has the knowledge of the different RSs affiliated to it. 
>>>> 
>>>>> (2) The client knows what kind of thing it’s looking for, but doesn’t know who to ask it from. 
>>>> Once again, the client could first talk to a "global service provider" which has the knowledge of the different RSs affiliated to it. 
>>>> 
>>>>> There’s a cross-domain trust that’s bridged by the AS, and the AS needs to dispatch to different resources depending on which user logged in (and possibly what the user consented to). To make things more concrete, the client needs to get driver’s license information, but doesn’t know ahead of time which of the many state/provincial bureaus to call to get that information because it doesn’t know yet who the user is. 
>>>> When a user has a driving license, he knows its issuer. The user can then provide some hint to the client.
>>>> 
>>>>> The AS will know who the user is once they log in and approve things, and so it can direct the client to call the correct RS. Since this is a relatively simple API with a pre-negotiated cross-domain trust, the AS returns a URL that the client presents the token at. 
>>>>  A single AS should not be aware of all the attributes a user has. 
>>>> 
>>>>> 
>>>>> As far as I know, in both of these cases, the token is only good for that API and not others — but more on that later.
>>>>> 
>>>>> A simple thing to do is just give back a URL with the access token, which tells the client where to go. 
>>>>> 
>>>>> 	{
>>>>> 		“access_token”: {
>>>>> 			“value”: “87yui843yfer”,
>>>>> 			“resource_uri”: “https://example/foo <https://example/foo>"
>>>>> 		}
>>>>> 	}
>>>>> 
>>>>> This is good for some kinds of APIs, but it’s limiting because not all APIs dispatch based on the URL. An AS might want to divvy up access tokens to an API that’s keyed on headers, or verbs, or any number of things. And it doesn’t tell us immediately what to do about non-exact URL matches. Can the client add query parameters and still use the token? What about path segments? I like that this simple approach addresses some common deployments that we already see today (see above), it’s not universal. Do we want or need a universal description language for directing where tokens go?
>>>>> 
>>>>> This also opens up a whole new set of security questions. If the AS can now direct the client where to use the token, could a rogue AS convince a legit client to use a stolen token at the wrong RS? And what if the client ignores the directions from the AS entirely? Could this open up new avenues of attack?
>>>>> 
>>>>> This is just the start, too. Things get even more complex if the client can ask for multiple different kinds of resources at once. What if the AS decides that the client needs a hyper-focused directed token for one part of the API, but can use a general token for other stuff? Can it signal that to the client? And if it can, does that mean that all clients need to be prepared for that kind of thing?
>>>>> 
>>>>> I firmly believe that whatever we build in GNAP, we need to optimize for the overwhelmingly common use case of a client getting a single access token to call APIs that it already knows about. Anything we add on top of that really can’t get in the way of this, because if it does, there’s very small chance that people will try to use this for everyday things. Keep the simple things simple, and the complex things possible, after all.
>>>>> 
>>>>> I’m really looking forward to hearing what the community thinks about these use cases, and hopefully the people I’ve chatted with offline about this can join the conversation and provide more light than I was able to.
>>>> The two use cases which are considered above are:
>>>> 
>>>> (1) The client knows what resource it’s calling, but it doesn’t know where it’s hosted.
>>>> (2) The client knows what kind of thing it’s looking for, but doesn’t know who to ask it from. 
>>>> 
>>>> That does not mean in any way that these two use cases should be solved by placing the AS at the centre of the solution.
>>>> 
>>>> Denis
>>>> 
>>>>> 
>>>>>  — Justin
>>>>> 
>>>>> 
>>>> 
>>>> -- 
>>>> Txauth mailing list
>>>> Txauth@ietf.org <mailto:Txauth@ietf.org>
>>>> https://www.ietf.org/mailman/listinfo/txauth <https://www.ietf.org/mailman/listinfo/txauth>
>> 
>> -- 
>> Txauth mailing list
>> Txauth@ietf.org <mailto:Txauth@ietf.org>
>> https://www.ietf.org/mailman/listinfo/txauth <https://www.ietf.org/mailman/listinfo/txauth>
>> 
>> This communication, including any attachments, is confidential. If you are not the intended recipient, you should not read it - please contact me immediately, destroy it, and do not copy or use any part of this communication or disclose anything about it. Thank you. Please note that this communication does not designate an information system for the purposes of the Electronic Transactions Act 2002.
> 
> -- 
> Txauth mailing list
> Txauth@ietf.org <mailto:Txauth@ietf.org>
> https://www.ietf.org/mailman/listinfo/txauth <https://www.ietf.org/mailman/listinfo/txauth>