Re: [OAUTH-WG] refresh tokens and client instances

[John Bradley <[e-mail]>]

I was assuming re-authenticating the user.   The thing is that with a public client you can't assume that it is the same instance of that client that requested the first grant for email, so you really should re-prompt for both scopes.

I have seen instances (mostly OAuth 1) where a AS is sloppy and remembers persistent grants( sometimes including default ones for the venders own app).  
An attacker gets the client_id and app_secret and impersonates the app.  If a user has pre-consented to scopes the attacker can by impersonating the legitimate app get scopes that the user has not consented to.

In OAuth 2 attack can take advantage of AS that don't require registering redirect URI for the code flow,  because they mistakenly confuse code with confidential.

In my opinion you should not use sticky grants with clients that are not confidential, and dynamic registration lets you make a native app confidential.
(no you can't guarantee the app is legitimate without some sort of out of band verification, but it is confidential)

John B.

On Jul 7, 2014, at 5:17 PM, Bill Mills <[e-mail]> wrote:

> I would argue that if the user approved a specific access profile that the token should be limited to that.  If it wants more rights/scope that should trigger an approval from the user based on a full authentication.  It's all a bit mudy though because you could imagine implementing a more limited scope the user would not have to approve.
> 
> 
> On Monday, July 7, 2014 2:09 PM, John Bradley <[e-mail]> wrote:
> 
> 
> Inline
> On Jul 7, 2014, at 4:59 PM, Sergey Beryozkin <[e-mail]> wrote:
> 
> > Hi John, All,
> > On 03/07/14 23:02, John Bradley wrote:
> >> Yes,
> >> 
> >> The the undifferentiated is initially differentiated by the user during Authorization by having a code returned and then by exchanging the code for a refresh token.
> >> It however returns to being undifferentiated on subsequent authorization requests.
> >> This makes having sticky grants (only asking for permission for incremental scopes) a potential security problem, as the AS has no way to know if the client is the one that the pervious authorization was intended for.
> >> 
> >> Some AS just assume that you want the same permissions across all instances of a client,  however if this is a public client then someone could impersonate the client app and basically do privilege escalation.
> >> 
> > Why would a public client holding a refresh token securely entered into it by a user request a new authorization without actually requesting the new scopes ? The client can just get a new access/refresh token from now on ?
> 
> A client holding a refresh token may want to add additional scopes, perhaps it only initially asked for permission to get a email address and now it wants a phone number.
> 
> If it is a public client the AS needs to ask for permission to grant both scopes,  it can't treat the email permission as sticky.
> > 
> >> What dynamic client registration gives us for native apps is a way to identify specific instances of clients at the authorization endpoint by having different client_id and validating that with instance specific client credentials.  This also prevents the use of code if it is intercepted in the reply from the authorization endpoint.
> >> 
> > Would it be fair to say that a dynamic client registration is a preferred method of registering *public* clients from now on, *unless*
> > no sticky grants are used in which case a typical/default registration mode is OK ?
> 
> It is up to the AS and how it wants to manage clients.  Some will not want to manage thousands of client_id, others won't mind.  
> 
> If you don't have sticky grants and can mitigate code being intercepted in the response by using http://tools.ietf.org/html/draft-sakimura-oauth-tcse ,
> then having a public client works.
> 
> 
> > 
> > Thanks, Sergey
> > 
> >> John B.
> >> 
> >> On Jul 3, 2014, at 12:28 PM, Sergey Beryozkin <[e-mail]> wrote:
> >> 
> >>> Hi
> >>> On 03/07/14 16:40, Bill Mills wrote:
> >>>> Implementations may/SHOULD bind refresh tokens to specific client
> >>>> instances.  Yes, it's possible that the client ID with dynamic
> >>>> registration is unique to each client, but many of the token theft use
> >>>> cases include the possibility of stealing the client ID too if you know
> >>>> you need to.
> >>>> 
> >>> What exactly is a 'client instance' when we talk about having a single client id registration, with the id shared between multiple devices (which is what I believe this thread started from).
> >>> 
> >>> What I understood, as far as the authorization service is concerned, a 'client instance' for AS is a combination of a client id + code grant.
> >>> 
> >>> + (optional) refresh token (as was mentioned earlier). But it appears to me a client instance can be uniquely identified by two values only without a refresh token.
> >>> 
> >>> When a user authorizes a given device and gets a grant code and enters it into the device securely we have a 'client instance' ready to get the tokens, with that device (client instance) using a client id and the grant code to get an access token and a refresh token.
> >>> 
> >>> Lets say it sends a "client_id=1&code=2" sequence to get the tokens.
> >>> A client id + a code value constitutes a client instance, because a code would be unique per every authorization, right ?
> >>> 
> >>> So the service will return an access token + refresh token to the device. Refresh Token could've been associated with a record containing a client id + grant code info earlier or at the moment the code is exchanged for an access token.
> >>> 
> >>> During the subsequent refresh token grant request we have "client id + refresh token" as a client instance.
> >>> 
> >>> I'm not sure what exactly I'd like to ask here :-), but I wonder if the above sounds right, then I guess I'd like to conclude that when we talk about the authorization code grant then a refresh token is not the only key that uniquely identifies a client instance:
> >>> 
> >>> Initially it is a client id + code grant, a refresh token does not offer an extra uniqueness at the point of the client device requesting an access token with a code. Refresh token only starts acting as the key client instance identifier at a refresh token grant time.
> >>> 
> >>> Sorry for a long email, I'm very likely missing something, so any clarifications will be welcome
> >>> 
> >>> Thanks, Sergey
> >>> 
> >>> 
> >>> 
> >>> 
> >>> 
> >>> 
> >>> 
> >>>> -bill
> >>>> 
> >>>> 
> >>>> On Thursday, July 3, 2014 4:33 AM, Sergey Beryozkin
> >>>> <[e-mail]> wrote:
> >>>> 
> >>>> 
> >>>> Hi
> >>>> 
> >>>> I'm finding the answers from John interesting but I'm failing to
> >>>> understand why refresh tokens are mentioned in scope of identifying the
> >>>> specific client instances.
> >>>> 
> >>>> AFAIK refresh tokens would only go on the wire, assuming they are
> >>>> supported, when a client exchanges a grant for a new access token.
> >>>> And when the client uses a refresh token grant.
> >>>> 
> >>>> Was it really about a refresh token grant where the incoming client id
> >>>> and refresh token pair can uniquely identify the actual client instance
> >>>> ? That would make sense.
> >>>> 
> >>>> Something else I'd like to clarify.
> >>>> John mentions a refresh token is created at the authorization grant
> >>>> initialization time. Would it make any difference, as far as the
> >>>> security properties of a grant are concerned, if refresh token was only
> >>>> created at a grant to access token exchange point of time ?
> >>>> 
> >>>> Thanks, Sergey
> >>>> 
> >>>> On 27/06/14 19:21, John Bradley wrote:
> >>>>> Inline
> >>>>> 
> >>>>> On Jun 27, 2014, at 1:24 PM, Madjid Nakhjiri <[e-mail]
> >>>> <mailto:[e-mail]>
> >>>>> <mailto:[e-mail] <mailto:[e-mail]>>> wrote:
> >>>>> 
> >>>>>> Hi John,
> >>>>>> Thank you for your reply. Would appreciate if you consider my inline
> >>>>>> comments below and respond again!
> >>>>>> R,
> >>>>>> Madjid
> >>>>>> *From:*John Bradley [mailto:[e-mail]
> >>>> <mailto:[e-mail]>]
> >>>>>> *Sent:*Wednesday, June 25, 2014 5:56 PM
> >>>>>> *To:*Madjid Nakhjiri
> >>>>>> *Cc:*[e-mail] <mailto:[e-mail]> <mailto:[e-mail]
> >>>> <mailto:[e-mail]>>
> >>>>>> *Subject:*Re: [OAUTH-WG] refresh tokens and client instances
> >>>>>> In 3.3 It is saying that the refresh token is a secret that the
> >>>>>> Authorization server has bound to the client_id, that the
> >>>>>> Authorization server effectively uses to differentiate between
> >>>>>> instances of that client_id.
> >>>>>> Madjid>>If I have 10,000s of devices, each with an instance of the
> >>>>>> OAUTH client, but they are all using the same client ID, how would the
> >>>>>> server know which token to use for what client? unless when I am
> >>>>>> creating a token, I also include something that uniquely identifies
> >>>>>> each instance? Don’t I have to use SOMETHING that is unique to that
> >>>>>> instance (user grant/ID?)?
> >>>>> When the grant is issued you create and store a refresh token which is
> >>>>> effectively the identifier for that instance/grant combination.
> >>>>> When it comes back on a request to the token endpoint you look up the
> >>>>> grants associated with it.  You also hack that the client_id sent in
> >>>>> the request matches to detect errors mostly)
> >>>>> 
> >>>>>> When the refresh token is generated, it can be stored in a table with
> >>>>>> the client_id and the information about the grant.  You could also do
> >>>>>> it statelesly by creating a signed object as the refresh token.
> >>>>>> Madjid>>agreed, but for the signed object to be self-sustained, again
> >>>>>> would I not need something beyond a “population” client_ID? Are we
> >>>>>> prescriptive what “information about the grant” is?
> >>>>> You would be creating a bearer token as long as the AS signs it you can
> >>>>> put whatever grant grant info you like in it, that is implementation
> >>>>> specific.  It  could be a list of the scopes granted and the subject.
> >>>>>> The spec is silent on the exact programming method that the
> >>>>>> Authorization server uses.
> >>>>>> Madjid>>Are there any other specs in IETF or elsewhere (OASIS, etc?)
> >>>>>> that prescribe token calculation (e.g. hash function, parameters, etc)?
> >>>>> 
> >>>>> You can look at JOSE and JWT for a way to create tokens
> >>>>> http://tools.ietf.org/html/draft-ietf-oauth-json-web-token
> >>>>>> In 3.7 Deployment independent describes using the same client_id
> >>>>>> across multiple instances of a native client, or multiple instances of
> >>>>>> a Java Script client running in a browsers with the same callback uri.
> >>>>>> Since the publishing of this RFC we have also developed a spec for
> >>>>>> dynamic client registration so it is possible to give every native
> >>>>>> client it's own client_id and secret making them confidential clients.
> >>>>>> Madjid>>I would need to look at those specs, however, I thought that
> >>>>>> the “confidential client” designation has to do with the client
> >>>>>> ability to hold secrets and perform a-by-server-acceptable
> >>>>>> authentication. Does dynamic client registration affect client’s
> >>>>>> ability in that aspect?
> >>>>> 
> >>>>> Yes it doesn't require the secret to be in the downloaded instance of
> >>>>> the native app.  It can be populated at first run, changing it from
> >>>>> public to confidential.
> >>>>> Confidential is not just for web servers any more.
> >>>>>> There is also a middle ground some people take by doing a proof of
> >>>>>> possession for code in native applications to prevent the interception
> >>>>>> of responses to the client by malicious applications on the device.
> >>>>>> https://datatracker.ietf.org/doc/draft-sakimura-oauth-tcse/
> >>>>>> John B.
> >>>>>> On Jun 25, 2014, at 8:06 PM, Madjid Nakhjiri <[e-mail]
> >>>> <mailto:[e-mail]>
> >>>>>> <mailto:[e-mail] <mailto:[e-mail]>>> wrote:
> >>>>>> 
> >>>>>> 
> >>>>>> Hi all,
> >>>>>> I am new to OAUTH list and OAUTH, so apologies if this is very
> >>>> off-topic.
> >>>>>> I am evaluating an OAUTH 2.0 implementation that is done based on bare
> >>>>>> bone base OAUTH2.0 RFC. From what I understand, many (or some) client
> >>>>>> implementations use a “global ID/secret” pair for all instances of the
> >>>>>> client.  Looking at RFC 6819 and there seem to be a whole page on this
> >>>>>> topic, if I understand it correctly. So questions:
> >>>>>> 1)Section 3.7 talks about deployment-independent versus deployment
> >>>>>> specific client IDs. I am guessing “deployment-independent” refers to
> >>>>>> what I called “global”, meaning if I have the same client with the
> >>>>>> same client ID installed in many end devices, that is a deployment
> >>>>>> independent case, correct?
> >>>>>> 2)Section 3.3 on refresh token mentions that the token is secret bound
> >>>>>> to the client ID and client instance. Could somebody please point me
> >>>>>> to where the token generation and binding is described? Also how is
> >>>>>> the client instance is identified?
> >>>>>> Thanks a lot in advance,
> >>>>>> Madjid Nakhjiri
> >>>>>> _______________________________________________
> >>>>>> OAuth mailing list
> >>>>>> [e-mail] <mailto:[e-mail]> <mailto:[e-mail]
> >>>> <mailto:[e-mail]>>
> >>>>>> https://www.ietf.org/mailman/listinfo/oauth
> >>>> 
> >>>>> 
> >>>>> 
> >>>>> 
> >>>>> _______________________________________________
> >>>>> OAuth mailing list
> >>>>> [e-mail] <mailto:[e-mail]>
> >>>>> https://www.ietf.org/mailman/listinfo/oauth
> >>>>> 
> >>>> 
> >>>> _______________________________________________
> >>>> OAuth mailing list
> >>>> [e-mail] <mailto:[e-mail]>
> >>>> https://www.ietf.org/mailman/listinfo/oauth
> >>>> 
> >>>> 
> >>> 
> >>> _______________________________________________
> >>> OAuth mailing list
> >>> [e-mail]
> >>> https://www.ietf.org/mailman/listinfo/oauth
> >> 
> > 
> 
>