Re: [GNAP] Should access tokens be opaque or not for the clients ?

[Justin Richer <[e-mail]>]

I think Steve’s use case gives an interesting example of something that I believe Adrian’s been talking about as well — delegating to another person.

We traditionally think of delegation at the AS to mean “a different user account logs in to the AS”, and it’s important to know that this does in fact work. But this could just as easily be “Steve sends his friend an artifact to present to the AS to get access”. This isn’t Steve (or Steve’s client software) getting and sending an access token, this is sending some other set of rights credential to his friend, or his friend’s software. So then Steve’s AS can have policy that says “when I see this rights credential, give an access token for the following attenuated things, because Steve said so”. The AS doesn’t even ever need to know who Steve’s friend is! 

And importantly here, we don’t expect Steve’s friend to bring their own AS to the party. Steve’s RS has no reason to trust his friend’s AS, let alone his friend’s client. But Steve’s AS is in a position to know more about the rights policies that Steve wants to apply, and so it can use other technologies like VC’s, DIDs, OIDC, or any number of other things to let things go through. This pattern is exactly why I think we need to think of this in terms of what’s currently called “interaction” at the AS. It’s not really just about how the end-user interacts, it’s about how the AS gathers all the consent and authorization information. I plan to rewrite section 4 to that effect.

 — Justin

> On Mar 18, 2021, at 1:58 PM, Stephen Moore <[e-mail]> wrote:
> 
> I'm having some concerns with the way you're flipping the trust here Denis. You seem to indicate that the user has trust in the RS and the Client and has to gain the trust of the AS. I feel like the least trustworthy portion of the architecture is the client, because that seems like it is the most interchangeable component.
>  Let me explain by way of a use case I have in mind.
> Let's say I have a smart home. I have devices and data storage and rule engines that are parts of sensors and hubs I buy for the house. I have a trust in those devices, because it's hardware, if the company is leaking data at the hardware level there isn't much I can do other than buy different sensors.
> In this case, the primary home hub, which is under control of the homeowner, would also be the AS, as well as a policy agent.  Authentication in the example is outsourced to whatever auth providers you want. I set up my home using a google account, a friend is staying and wants access and uses FB for auth... whatever.
> The part that I have little control over is the client my friend uses on his phone. Or if I change the client on the tablet in the kitchen to try out something new. These pieces might be single page web apps, they might be downloaded apps, etc. I do not want them having any additional information than is necessary. If I give my friend permission to unlock the door, and control the lights because they are feeding my cats while I'm away, the client doesn't need to know anything beyond, I see this home controller, and I have access to these lights and that door lock.
> 
> In your scenarios, the client is spy by design, and if it is a web app, I don't trust them to not grab my user email address from the AS and put it on a mailing list it sells to some sketchy company, when the only other component that needs that information is the RS.
> 
> -steve
> 
> On Thu, Mar 18, 2021 at 1:21 PM Denis <[e-mail] <mailto:[e-mail]>> wrote:
> I have changed the title of this thread, since the original topic has been closed. 
> 
> Justin, you are raising the following question:
> The question I have is — why does the client need to know what’s in the token? 
> Not if they could possibly know, but why would we expect a client to know and manage the contents of the token? 
> The answer is quite simple: so that the end-user may have confidence into the architecture. 
> The seventh of the eleven privacy principles from ISO 29100 (page 14) is:
> 7. Openness, transparency and notice 
> If the access token is considered as a black box for the end-user, it cannot be confident about the content of the access token.
> Such access token may disclose some private information without the consent of the end-user. If the access token that it has obtained 
> does not match with what has been requested, the client will first not forward the access token to the RS and then will complain about the AS.
> 
> Denis
> 
> 
>> Hi Justin, 
>> 
>> "The access token is fundamentally a conversational artifact between the AS and the RS which the client is the carrier of."
>> That's right, and a good opportunity to distinguish between the need for token inspection (i.e. a runtime verification, wherever that may be) 
>> and the need to negotiate the token format (a discovery). That last part could very well be an extension, as a way to open up the ecosystem 
>> beyond JWTs (which are fine, but come with their own limitations). 
>> 
>> Fabien
>> 
>> On Thu, Mar 18, 2021 at 5:59 PM Justin Richer <[e-mail] <mailto:[e-mail]>> wrote:
>> On Mar 18, 2021, at 11:30 AM, Alan Karp <[e-mail] <mailto:[e-mail]>> wrote:
>>> 
>>> Justin Richer <[e-mail] <mailto:[e-mail]>> wrote:
>>> On Mar 18, 2021, at 12:08 AM, Alan Karp <[e-mail] <mailto:[e-mail]>> wrote:
>>>> 
>>>> Adrian Gropper <[e-mail] <mailto:[e-mail]>> wrote:
>>>> 
>>>> Is there an AS involved in the delegation? How and where in the lifecycle of the protected resource? 
>>>> 
>>>> If tokens are certificates, the AS need not be involved in subsequent delegations.  The AS must be involved if the tokens are opaque. 
>>> 
>>> Tokens are opaque to the client instance. They are not opaque to the AS. They might be opaque to the RS, but that depends on the kind of relationship the RS and AS have. GNAP should allow different options here as there are different use cases for that.
>>> 
>>> Tokens are not opaque to the client in SPKI, zcap-ld, Orie's implementation with VCs, or our Zebra Copy work.  Why must they be in GNAP?
>> 
>> The existence of the AS is exactly the reason for this. The AS is the role that “knowledge about the token contents” has been outsourced to in the GNAP model (which is based on the OAuth model).
>> 
>> It brings significant simplicity for the client developer. The question I have is — why does the client need to know what’s in the token? Not if they could possibly know, but why would we expect a client to know and manage the contents of the token? 
>> 
>> The access token is fundamentally a conversational artifact between the AS and the RS which the client is the carrier of. The client is not the audience of the token, nor the creator of the token, nor even the manager of the token and the rights it represents. The client as a simple carrier is a powerful model that allows the security layer to get out of the way of the actual application logic that developers want to do.
>> 
>>  — Justin
>> 
>>> 
>>> --------------
>>> Alan Karp
>>> 
>>> 
>>> On Thu, Mar 18, 2021 at 4:56 AM Justin Richer <[e-mail] <mailto:[e-mail]>> wrote:
>>> On Mar 18, 2021, at 12:08 AM, Alan Karp <[e-mail] <mailto:[e-mail]>> wrote:
>>>> 
>>>> Adrian Gropper <[e-mail] <mailto:[e-mail]>> wrote:
>>>> 
>>>> Is there an AS involved in the delegation? How and where in the lifecycle of the protected resource? 
>>>> 
>>>> If tokens are certificates, the AS need not be involved in subsequent delegations.  The AS must be involved if the tokens are opaque. 
>>> 
>>> Tokens are opaque to the client instance. They are not opaque to the AS. They might be opaque to the RS, but that depends on the kind of relationship the RS and AS have. GNAP should allow different options here as there are different use cases for that.
>>> 
>>> It would probably be worthwhile to separate the portions of the spec that talk about the RS-AS relationship into its own standalone document. A similar approach was taken in UMA2 and it was helpful. (Though admittedly, as with anything, there are missteps there that we can hopefully learn from.)
>>> 
>>>  — Justin
>>> 
>>>> 
>>>> --------------
>>>> Alan Karp
>>>> 
>>>> 
>>>> On Wed, Mar 17, 2021 at 8:54 PM Adrian Gropper <[e-mail] <mailto:[e-mail]>> wrote:
>>>> Sure! 
>>>> 
>>>> Is there an AS involved in the delegation? How and where in the lifecycle of the protected resource? 
>>>> 
>>>> Also your use of "the client" seems to imply that either there is only one client or the client doesn't matter. Which is it?
>>>> 
>>>> Adrian
>>>> 
>>>> On Wed, Mar 17, 2021 at 11:43 PM Fabien Imbault <[e-mail] <mailto:[e-mail]>> wrote:
>>>> Thanks for that. 
>>>> 
>>>> Trying to reframe it:
>>>> GNAP is defined as a delegation protocol so the main intent is related to a delegate of the RO (i.e. the end user) that wishes to access the RO's protected resources, through the client. 
>>>> 
>>>> Fabien 
>>>> 
>>>> Le jeu. 18 mars 2021 à 04:29, Adrian Gropper <[e-mail] <mailto:[e-mail]>> a écrit :
>>>> At various points in the lifecycle of the protected resource the client at the resource server (RS) might be:
>>>> The RO (subject) user agent trading payment for a service promise
>>>> The RO user agent using the promise to access the protected resource
>>>> A delegate of the RO user agent using a different client
>>>> What's vague is where the GNAP AS enters the picture as described above. How would you describe it?
>>>> 
>>>> Adrian
>>>> 
>>>> 
>>>> On Wed, Mar 17, 2021 at 10:20 PM Fabien Imbault <[e-mail] <mailto:[e-mail]>> wrote:
>>>> Hi Adrian
>>>> 
>>>> I'm still confused why you're saying the terminology is vague. 
>>>> I get the "power" neutrality is not to your liking, but RQ / user agent is no better in my view. 
>>>> 
>>>> Can you elaborate? 
>>>> 
>>>> Fabien 
>>>> 
>>>> Le jeu. 18 mars 2021 à 00:18, Adrian Gropper <[e-mail] <mailto:[e-mail]>> a écrit :
>>>> I'm sure you're right. Our vague terminology around client and end-user leads to my confusion. If GNAP is primarily about delegation then, of course, we should avoid any incentives to impersonate or we're wasting our time. This is partly why I'm trying to study up on capabilities and asking for expert advice from folks like Alan Karp and Mark Miller (cc'd)
>>>> 
>>>> As best I can understand it, the RS has only two choices, it can:
>>>> store an attribute of the RO a [DID, email address, GNAP AS URL], or
>>>> hand the RO a capability as a sort-of promise and avoid making any entries in an ACL or equivalent.
>>>> When a token comes back to the RS, it will either be:
>>>> validated according to something associated with the stored RO attribute, or
>>>> signed by the RS itself.
>>>> Either way, trust in the client seems moot.
>>>> 
>>>> Adrian
>>>> 
>>>> 
>>>> 
>>>> On Wed, Mar 17, 2021 at 5:29 PM Justin Richer <[e-mail] <mailto:[e-mail]>> wrote:
>>>> On Mar 17, 2021, at 4:55 PM, Adrian Gropper <[e-mail] <mailto:[e-mail]>> wrote:
>>>>> 
>>>>> On Wed, Mar 17, 2021 at 4:23 PM Tobias Looker <[e-mail] <mailto:[e-mail]>> wrote:
>>>>> <snip>
>>>>> > A client might not have a DID but it could have a VC as a certificate of authenticity linked to some audit mechanism.
>>>>> 
>>>>> To me a VC would come under the assertions umbrella (that is to say a VC could be one type of valid assertion). The client may possess or been presented with a VC that it could include in its request to the AS as a way to identify the subject and perhaps prove authentication and authorization.
>>>>> 
>>>>> I do not assume that the client that interacts with the AS to make a request and receive a token is the same as the client that will present the token to the RS. In the US HIPAA use-case, for example, the root of trust is a contract between the patient-subject and the doctor-requesting party but the doctor workflow is expected to delegate the token to some                                                           other end-user that may be using a totally different client such as an EHR.
>>>>> 
>>>> 
>>>> If the client that gets the token is not same as the client that uses the token, that is a violation of core security principles as it allows for (and really designs for) impersonation by client software. I would have no reason to trust client software that would hand its credentials over to another piece of software, and in fact I shouldn’t trust it. 
>>>> 
>>>> I think you may be conflating several different kinds of parties under the “client” umbrella here, though. It’s entirely possible that one client might call an RS that in turn acts as a client for something else down stream. But each of those hops is different from the last.
>>>> 
>>>>  — Justin
>>>> 
>>> 
>> 
>> 
> 
> -- 
> TXAuth mailing list
> [e-mail] <mailto:[e-mail]>
> https://www.ietf.org/mailman/listinfo/txauth <https://www.ietf.org/mailman/listinfo/txauth>
> -- 
> TXAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/txauth