Re: [OAUTH-WG] New Version Notification for draft-lodderstedt-oauth-jwt-introspection-response-00.txt

Torsten Lodderstedt <> Sun, 27 May 2018 15:25 UTC

Return-Path: <>
Received: from localhost (localhost []) by (Postfix) with ESMTP id 500FE12EBC8 for <>; Sun, 27 May 2018 08:25:25 -0700 (PDT)
X-Virus-Scanned: amavisd-new at
X-Spam-Flag: NO
X-Spam-Score: -2.621
X-Spam-Status: No, score=-2.621 tagged_above=-999 required=5 tests=[BAYES_00=-1.9, RCVD_IN_DNSWL_LOW=-0.7, RCVD_IN_MSPIKE_H3=-0.01, RCVD_IN_MSPIKE_WL=-0.01, SPF_PASS=-0.001] autolearn=ham autolearn_force=no
Received: from ([]) by localhost ( []) (amavisd-new, port 10024) with ESMTP id 9sdFivKMjuHH for <>; Sun, 27 May 2018 08:25:21 -0700 (PDT)
Received: from ( []) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by (Postfix) with ESMTPS id 86B9812EB1B for <>; Sun, 27 May 2018 08:25:21 -0700 (PDT)
Received: from [] (helo=[]) by with esmtpsa (TLSv1.2:ECDHE-RSA-AES256-GCM-SHA384:256) (Exim 4.90_1) (envelope-from <>) id 1fMxXq-0000n2-He; Sun, 27 May 2018 17:25:18 +0200
From: Torsten Lodderstedt <>
Message-Id: <>
Content-Type: multipart/signed; boundary="Apple-Mail=_C63055C7-A910-4B41-AD87-649748BE370D"; protocol="application/pkcs7-signature"; micalg=sha1
Mime-Version: 1.0 (Mac OS X Mail 11.3 \(3445.6.18\))
Date: Sun, 27 May 2018 17:25:16 +0200
In-Reply-To: <>
Cc: oauth <>
To: Neil Madden <>
References: <> <> <> <> <>
X-Mailer: Apple Mail (2.3445.6.18)
X-Df-Sender: dG9yc3RlbkBsb2RkZXJzdGVkdC5uZXQ=
Archived-At: <>
Subject: Re: [OAUTH-WG] New Version Notification for draft-lodderstedt-oauth-jwt-introspection-response-00.txt
X-Mailman-Version: 2.1.22
Precedence: list
List-Id: OAUTH WG <>
List-Unsubscribe: <>, <>
List-Archive: <>
List-Post: <>
List-Help: <>
List-Subscribe: <>, <>
X-List-Received-Date: Sun, 27 May 2018 15:25:26 -0000

Hi Neil,

> Am 25.05.2018 um 12:49 schrieb Neil Madden <>om>:
> Hi Torsten,
>> On 25 May 2018, at 10:35, Torsten Lodderstedt <> wrote:
>> Hi Neil,
>>> Am 28.03.2018 um 17:41 schrieb Neil Madden <>om>:
>>> I like this draft, but I want to clarify if it is intended that the response JWT could be interpreted as an OpenID Connect ID Token? As the set of claims can overlap (in particular, all required ID token claims are valid token introspection response fields) and it seems highly likely that an AS will use the same keys for signing both (and it definitely will when the client_secret is used for signing), the signed response JWT could well be indistinguishable from an ID token (for the resource owner) with some additional claims.
>> Conceptually, the introspection response, an ID Token and even structured access tokens are quite the same if they carry user identifiers or claims. They identify the respective user account towards RP or RS and provide additional attributes, which, for example, can be used to meet access control decisions. And if JWT is the representation, they are also syntactically (nearly) equivalent. The main difference I see that an ID Token may contain a nonce, which is not required in access tokens.
> Are they the same? An ID token authenticates (claims about) a user to the RP.

The same definition holds true for access tokens towards RSs.

> I don’t think that is the intended interpretation of a token introspection response, and it’s not the purpose of an access token. My point is that as specified they can be syntactically identical and yet intended for different purposes.

The difference is just the role the party is taking in the overall protocol flow. In my experience, contents of access tokens and contents of ID tokens tend to be rather similar, at least in the consumer identity space. 

>>> If this is not the case, then maybe consider adding a “crit”: [“scope”] claim to the response ( to indicate that the scope claim must be understood.
>> What do you want to achieve/prevent?
> The sorts of issues discussed in
> As I said in a follow-up email, “crit” does not actually work here as it only applies to headers. If we wanted to distinguish them, we could use explicit typing as recommended in the BCP: and have a type of “introspection+jwt” or something like that (with an appropriate media type registration).

The text also states "Note that this is a specific type of substitution attack.“ and I fully agree with this assessment. In my opinion, the consequence is that all the countermeasures available for token substitution can and must be used to detect such misuse. In the particular case of an an attacker wanting to use a token introspection response JWT as ID token in a different flow, all security checks defined by OpenId Connect ( must be performed. This means the RP must, for example, validate issuer, audience and nonce. This should be sufficient to prevent any kind of abuse.

>>> I can think of one potential use-case (I’ll let you decide the merits of it) where it might actually be useful to explicitly allow the response to be an ID Token. Consider an application (RS) that uses a traditional authorization model: it authenticates a user, sets a cookie, and then based on who that user is makes dynamic access control decisions to see what they are allowed to do (e.g., ACLs, RBAC, whatever). An easy way to upgrade this app to modern standards would be to replace the home-spun authentication system with OIDC, but leave the rest in place. Now the system uses OIDC to authenticate the user, sets the ID token as the cookie, and then still applies the same access control decisions that it always has done.
>>> Now imagine that a new requirement comes in to support OAuth 2.0 access tokens to allow delegation to third-party apps. A really simple way to achieve that would be to put a filter/reverse proxy in front of the RS that extracts access tokens coming in, performs signed JWT token introspection against the AS to validate the token and then checks the the scopes are appropriate for the request. It can then simply replace the access token in the original request with the signed token introspection response (as ID token) and forward it on to the original RS server. As the introspection response is a valid ID token for the resource owner, the RS will then apply all its normal access control checks to ensure that the resource owner actually has the permissions that they have delegated to the client.
>>> I think potentially that is quite an interesting application of this draft, but I don’t think it was intended! I think probably a decision should be made as to whether that kind of usage should be allowed and explicitly adjust the draft to either allow or deny it. If it is allowed, then possibly there should be a way for the caller to hint that they want the response to be a valid ID token.
>> I don’t currently see a way the introspection response could be abused as ID Token other than the recipient of the response (or an attacker able to obtain the response object) sets up a fake IDP and provides the response as an ID token to a RP. The RP should be able to detect this (as other ways to replay ID tokens) by utilizing the nonce.
> As a client I can go and get an access token with some scope “a”. I now hit the token introspection endpoint and get back a signed JWT. That signed JWT happens to be syntactically identical to an OIDC ID Token and the AS happens to use the same key to sign both. So now I effectively have an ID Token for that user, although I never requested the “openid” scope. OK, so I don’t have an access token with the right scopes to hit the UserInfo endpoint, and I learn nothing that I could not get from a normal introspection response, but still it feels a bit surprising to me.
> Furthermore, as the “aud” claim in the token introspection response is optional, and it has no nonce, I *might* be able to replay this JWT in other situations to other parties. For instance, if I can find an OIDC RP that does not use or validate nonces correctly (they are optional), then I may be able to use it to log in as the user at that other RP.

Good point. I think that would work for an RP using the id_token (or a hybrid) response type. But the RP MUST validate the aud claim.

> I don’t think these are huge issues as they depend on RPs already being careless about validation (e.g., ignoring a missing audience claim), but they are the sort of non-obvious interactions between components where security issues often crop up, so I want to ensure the possibility has been considered by the WG and an informed decision made either way.

I fully agree. I would like to prevent a situation where somebody believes we need to adopt a different syntax in order to prevent such "Cross-JWT Confusion“ attacks. 

> I would be happy if we just made the “aud” and “iss” claims mandatory in this spec (they are optional in RFC 7662), and require that the protected resource is in the audience.

Makes sense. I will add this to the draft. 


> — Neil