Re: [OAUTH-WG] Fixing the Authorization Server Mix-Up: Call for Adoption

John Bradley <> Tue, 23 February 2016 23:17 UTC

Return-Path: <>
Received: from localhost ( []) by (Postfix) with ESMTP id 191201B3763 for <>; Tue, 23 Feb 2016 15:17:23 -0800 (PST)
X-Virus-Scanned: amavisd-new at
X-Spam-Flag: NO
X-Spam-Score: -1.901
X-Spam-Status: No, score=-1.901 tagged_above=-999 required=5 tests=[BAYES_00=-1.9, DKIM_SIGNED=0.1, DKIM_VALID=-0.1, SPF_PASS=-0.001] autolearn=ham
Received: from ([]) by localhost ( []) (amavisd-new, port 10024) with ESMTP id hm-WmNncDK_s for <>; Tue, 23 Feb 2016 15:17:19 -0800 (PST)
Received: from ( [IPv6:2a00:1450:400c:c09::231]) (using TLSv1.2 with cipher ECDHE-RSA-AES128-GCM-SHA256 (128/128 bits)) (No client certificate requested) by (Postfix) with ESMTPS id A02261B3759 for <>; Tue, 23 Feb 2016 15:17:18 -0800 (PST)
Received: by with SMTP id c200so244564803wme.0 for <>; Tue, 23 Feb 2016 15:17:18 -0800 (PST)
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed;; s=20150623; h=mime-version:subject:from:in-reply-to:date:cc:message-id:references :to; bh=BY/vosOYFD3DwKjvq9RoyNSdyN+fSJTpg0R/RRnAI/U=; b=IUkVnUiIqQJjrju5hQGf2cirhGxw2jQyg3vdp3C0GrmKuhnjbb98+gijhYKn/+D9p8 5BZbjMezA5XDx8x/3c840CMS+7BtfP0xqowInCkKKIxfu2VYYGnuh9L/Pd91Jfxf5guv Yzcqg3bLrnHKcTTz5XPTu4uHwOeyiib+NNBUXX1tC/n8MSR22fCYixtEE84UgPFnbjQn bGi1xVas/Mu/1Qo04BYmx5fKx42qk7TmQ22pgDiRfiIz1zy4sqygK1v+ZJf7bUPxz1FN lKjvhZyEj5J3WhNFKrMaoV9xGl8W3GqmTI+RtJOZQek5HdQ+NKRqGHJrlPvipevwKZ13 +6Nw==
X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed;; s=20130820; h=x-gm-message-state:mime-version:subject:from:in-reply-to:date:cc :message-id:references:to; bh=BY/vosOYFD3DwKjvq9RoyNSdyN+fSJTpg0R/RRnAI/U=; b=aCm8as4iM8kbJTyDY1I4NqbhgkQsqHkLBMN9aMbjKxt6qh4QRdnMi6grq99X/jd3JH qVeSQrSqy5X16WmooR11AHIrny5WiHRqTcJ0A4D0mipNTBeS5TfSVDcZmD9VyYbyKg18 5L/WX+qW0a9VytRou6DvoHvTXiAbCkBS00WgMWon7/t+is+CFdI3FVSeYcWDcG1way+k 8ImgKz+f8OoyOf4dPBgZZCvvvKttd7Apsk1cYjqgZxDAKMzbyD0J2zUjfvF7GA70mkgD M/davYw/1uGsQiTyFCJe62pIV+5Is6kwx/s9a2veb1suYXj0h7kCKG8RjgvWoLBvJXFq ajwA==
X-Gm-Message-State: AG10YORTBoavB2OrQI2RH5wgH/F/mOUcscPE/r0E3QEn5Dt/X37xLTBTnbFNBf0XURiOwg==
X-Received: by with SMTP id a5mr21984685wmd.32.1456269437063; Tue, 23 Feb 2016 15:17:17 -0800 (PST)
Received: from [] ( []) by with ESMTPSA id cb2sm117845wjc.16.2016. (version=TLS1 cipher=ECDHE-RSA-AES128-SHA bits=128/128); Tue, 23 Feb 2016 15:17:15 -0800 (PST)
Content-Type: multipart/signed; boundary="Apple-Mail=_A86BEBCF-DB8F-4962-AB37-96C4E2AF87DA"; protocol="application/pkcs7-signature"; micalg="sha1"
Mime-Version: 1.0 (Mac OS X Mail 9.2 \(3112\))
From: John Bradley <>
In-Reply-To: <>
Date: Wed, 24 Feb 2016 00:17:12 +0100
Message-Id: <>
References: <> <> <> <> <> <> <05af01d16d4a$20230af0$606920d0$> <> <>
To: Roland Hedberg <>
X-Mailer: Apple Mail (2.3112)
Archived-At: <>
Cc: "<>" <>
Subject: Re: [OAUTH-WG] Fixing the Authorization Server Mix-Up: Call for Adoption
X-Mailman-Version: 2.1.15
Precedence: list
List-Id: OAUTH WG <>
List-Unsubscribe: <>, <>
List-Archive: <>
List-Post: <>
List-Help: <>
List-Subscribe: <>, <>
X-List-Received-Date: Tue, 23 Feb 2016 23:17:23 -0000

The idea of including the state in the PKCE calculation was one I think Hans brought up at the meeting in Darmstadt.

I think we discounted it as not solving the problem because the attacker could manipulate the code_challenge.

As and example the attacker receives the request and changes the code_challenge to one it makes up based on the state from a second request made to the client in it’s browser. 

The attacker can replay the token to the client with the state that will validate the code_callenge or use the token directly at the token endpoint to get a access token as it would know both the client secret and code verifier.

We moved on to look at another way to validate state by passing it to the token endpoint for validation to stop the cut and paste attack.

In thinking again about William and Nat’s question about why not PKCE, I looked at the idea of using PKCE again.

The problem with using PKCE as it is now is that we made the assumption that the attacker could not modify the request.

That got me trying to think of a way to:
 a) stop the attacker from modifying the request (signed request, probably too complicated to be generally deployed)
 b) allow the client to detect if code_challenge has been modified and abort of the request was compromise.

I realized that you could simply echo back the code_challenge from the authorization endpoint and that would allow the client to know if code_challenge had been altered in the request, assuming the attacker can’t modify the response from the legitimate AS.

So if we add that step we now have a client where code_challenge can’t be altered.  

That in it selfs stops the cut and pate attack.  It however still allows an attacker to use the code it receives along with the the code_verifier and client secret to get a AT and RT from the legitimate token endpoint.

So that is not enough Including state in the hash calculation helps against the cut and paste attack but we fixed that by returning code_challenge.

That leaves the other things we were talking about returning from the authorization endpoint.

If we include token_endpoint URI in the hash then any token the client requests thinking it is for the attackers token endpoint won’t generate the same code_challenge as the AS would based on it’s token endpoint.   The attacker would know the code_challenge and the real token endpoint URI but would have to generate a collision in the lifetime of code to use the code.   This would not be possible based on the PKCE analysis.

Returning the code_challenge and including the token_endpoint URI in the hash is the minimum needed for the code flow to stop both cut and paste as well  as protect the code from being used if stolen via a confused client attack.   It doesn’t stop the theft just the use of the token.  So it is mitigating the issue a different way from A and B that are trying to stop the attacker from getting the code.

Given the technique I realized it would be simple enough to also include some other paramaters in the hash
1) The authorization endpoint in the hash causing it not to validate if the client has been given a bad authorization endpoint so the client can modify the request.  
2) The client_id so that it cannot be tampered with in the request.
3) The state  This is tied to the browser instance for XSRF protection.  Including this means it cannot be changed and is another protection against cut and paste.

Including all 4 in the hash may be overkill as they overlap in protection.  I however think that not integrity protecting one of the values might bite us in the future.

We could go farther and include scopes or the whole request but that gets more complicated than it is worth.  We do have a option for request signing already.

So we almost had this at the Darmstadt meeting but it slipped away.

This is not stoping the attack it is protecting code.   

I should point out that some of the attacks like man in the middling registration can still compromise the client secret, making it possible for an attacker to impersonate a client.

Nat’s option B doesn’t have a solution to that ether as far as I can tell.

I think the only option for that is to have a logical identifier for the AS and some sort of discovery check at registration to be certain that you are using the correct registration location.

Option A plus discovery works for that and mitigates confused client but not cut and paste.

We still have a problem with AT leaking.   I think that needs to be dealt with separately. 
Access tokens should have a audience (by value or by introspection) the client needs to tell the AS what resource it want s to use the token at and have that included as the audience or the request rejected because it is an invalid audience.

That should happen at the token endpoint for code.
For implicit it would need to be at the authoriation endpoint and the audience would need to be echoed back to prevent tampering.

The AT can also be protected by POP so I think that code and AT should be kept separate issues.

John B.

> On Feb 23, 2016, at 9:59 PM, Roland Hedberg <> wrote:
> In line !
>> 22 feb 2016 kl. 05:08 skrev John Bradley <>:
>>> On Feb 22, 2016, at 9:22 AM, Nat Sakimura <> wrote:
>>> The risk impact of [case2] is more OAuth specific. The token is stolen as the token is going to be sent to a rogue resource, and the resource can use that to obtain the resource that was supposed to be only available to the proper client.
>>> The risk impact of [case3] is the most grave among these three. Now the attacker can create his own client that impersonates the compromised client.
>>> OAuth-mix-up
>>> ---------------------------
>>> OAuth-mix-up draft gives one way of addressing [case2] by introducing a new concept of “issuer” and “discovery” to RFC6749. It also returns client_id and verifies it in 4.2, but if the BadAS has assigned a same client_id to the client, it does not help.
>> Client_id are not guaranteed to be unique.  They need to be name spaced by the AS.   In OAuth currently we have no name for the AS this makes that difficult, the client can use the authorization endpoint URI, but that logic may well break in multi tenant situations.   Having a clear issuer string makes it easier for the client.
> I stumbled over exactly this point when I made my implementation follow the Mix-Up draft.
> If not discovery is involved I think an AS has to have a name which is different from the endpoint URIs.
>>> One of the issue with this approach is that the central notion of “issuer” is new to the existing servers and clients. The verification rule in 4.1 states “Compare the issuer URL for the authorization server that the client received when it registered at the authorization server”, but in most existing pure OAuth cases, there is no such thing, so you cannot compare. This means that you would have to re-register, and if you are doing that, the per-AS redirect_uri seems to be a much simpler solution.
>> The client developers I have talked to really hate per AS redirect URI as being too awkward for deployments.
> I on the other hand found it quite easy to do per AS redirect URIs, so it might well be implementation
> dependent rather then contextual.
>>> Per AS Redirect URI
>>> --------------------------------
>>> This does not involve wire protocol changes. It just adds requirements on the redirect uri. This by far is the simplest solution for [case2] (and [case1]), IMHO.
>>> Again, it is not a general framework for finding out tainted communication, so it may have other holes.
>> This is probably the hardest for the client developer and for the deployer.  Yes it is simplest from a spec point of view.
>> We need more developer feedback on this.
> As I said above I found this quite simple to implement.
>>> (Extended) PKCE
>>> ---------------------------------
>>> To begin with, it works only with code flow. There has to be something else to deal with implicit flow.
>>> PKCE binds the authorization request/response to the token request.
>>> If used with a confidential client, it seems to mitigate the vulnerability.
>>> John points out that it is not the case. I must be missing something here… but my logic goes like:
>>> 1.     The good client creates code_challenge-v and code_verifier-v=S256(code_challenge-v) and sends the client_id-v + code_challenge-v + state to the BadAS Authz EP.
>>> 2.     BadAS as a client prepares a code_verifier-b and code_challenge-b=S256(code_verifer-b).
>>> 3.     BadAS redirects to GoodAS with the client_id-v + code_challenge-b + state-v.
>>> 4.     GoodAS stores the code_challenge-b.
>>> 5.     GoodAS returns code-v + state-v to the client’s redirect uri.
>>> 6.     The client finds the AS from the state and sends code-v + code_verifier-v + secret-b to the BadAS token endpoint. Now, code-v and code_verifer-v is phished.
>>> Now the attacker tries to use the code-v and code_verifier-v.
>>> ### Case A:
>>> 7.     The BadAS as a client sends client_id-v + … but he does not have client secret for the good client, so it fails.
>>> ### Case B:
>>> 8.     The BadAS as a client sends client_id-b + code-v + code_verifier-b + secret-b etc. to GoodAS.
>>> 9.     GoodAS verifies the code_verifier-b is associated with code-v, but that code-v is not associated with client_id-b, so the token request fails.
>>> ### Case C: cut-n-paste
>>> 10.  The attacker launches cut-n-paste attack by replacing the code-b with code-v.
>>> 11.  The verifiers does not match, so fails.
>>> Please let me know what I am missing.
>> In a step 0 the attacker has the good client create another request in the attackers user agent to get state-0 and code_challange-0
>> Step 2 is not required.
>> Step 3 Bad AS redirects to good AS with client_id_v + state-v + code_challenge-0
>> Step 4 GoodAS stores code_challenge-0
>> Step 5 GoodAS returns code-v + state-v to the clients redirect_uri
>> Step 6 The client finds the AS from the state and sends code-v + code_verifier-v + secret-b to the BadAS token endpoint. Now, code-v and code_verifer-v is phished.
>> Case C1 :  cut and paste
>> 10.  The attacker launches cut-n-paste attack by inserting code-v into a response using state-0
>> 11. The client sends code-v and based on state-0 it sends code_verifyer-0 to the good AS token endpoint.
>> 12. The GoodAS verifies that code-verifyer-0 is correct for code_challange-0 that it bound to code in step 4
>> 13. The GoodAS receives RT + AT.
>> 14. The attacker has now used the client to bind the users resource to it’s account and is transferring money or looking at your data.
>> This could be 3rd party financial app like Mint as an example or photos or any other PII that could then be used to escalate identity theft.
>> This variation of the attack combining cut and paste with confused client was not mentioned in the research papers.
>> I found it looking to see if PKCE could be used to mitigate the confused client attack.
>> As I mentioned in a response to William, while the current PKCE Challenge methods only make the attack harder by forcing the attacker to get the client to make a step 0 request to get a code_challenge to use in the request,  we could define a new challenge method that would be effective.
>> That would remove the need to have a separate mechanism to prevent cut and paste.
>> The problem is that the PKCE challenge is independent of state so becomes vulnerable to the confused client.
>> What we would need to do is include state in the challenge so that if the AS receives mismatched state and code_challange in step 3 the verification of code verifier will fail. Effectively combining the cut and paste mitigation with PKCE.
>> I haven’t had time to write this up, but if the code_challenge == SHA256 (SHA256(state) || token_endpoint URI || Authorization_endpoint URI || client_id || code-veriyer) ,  then the attacker could not change state, client_id, or token_endpoint  independently of code_challenge.  (note I am using a hash of state to make storage size deterministic for the AS on the grounds that the client already needs to be able to do the hash) (Some attacks  change the client_id in the request so I am including it in the hash)
>> This is slightly more complicated than than S256, but not that much.  I wish I had thought of this when we were doing PKCE.   Hit me with the stupid stick, my fault.
>> I don’t know if it stops all the confused client attacks though.
>> If the client is confidential then it gives up it’s client secret assuming compromised registration, so we can’t really count on that for symmetric client authentication.
>> By including the token endpoint in the comparison if the client is sending the code to the attacker the client will use a different token endpoint_uri in to calculate the code_challenge than the GoodAS will use to verify it.    This would stop both cut and paste as well as stopping the attacker from using the code if they get it.   The attacker can’t get Secret for state-0, so it can’t create code_challenge that would be valid.
>> This stops the registration attack where the client gets a bad AS discovery endpoint that has all the Good AS info including dynamic registration endpoint but the bad AS token endpoint.   The bad AS would get the token, client_secret and code_verier, but will fail pkce verification because the token endpoint will be wrong.
>> The attack where the client registers with the good AS but with bad token endpoint and Authorization endpoint gives the attacker the ability to change the code_challenge that the Good AS is going to see.   It would need to make a new challenge using state and the GoodAS token endpoint and it’s client_id.
>> To do that it needs the code_verifier to use to calculate the hash to use as the code_challenge value.  As long as the client is not tricked into accepting a replay of the authentication response we should be safe.  The client would need to do replay protection on the code_verifier values before it makes a request to the token endpoint.
>> The BadAS could however make up a new code_challenge and code_verifier and use that in the request. For this one the AS would need to do pull discovery on the client to get a key, or the AS needs to return something in the response.  This attack can completely proxy all the endpoints as far as the client is concerned and is taking advantage of the AS saying you are granting permission to site X based on the redirect URI.
>> I can’t see PKCE on it’s own being able to stop a client from being used as a redirector back to the attacker unless you also returned the code_challenge in the response from the authorization endpoint.
>> Hypothetically if we returned code_challenge in the response from the authorization endpoint and have a new PKCE challenge method we might find it covers all of the attacks.
>> We would need to put some serious analysis into this to see if it really covers all the attacks.
>> It however doesn’t address the “token” response_type or steeling the access token by impersonating the RS.
>> I think the correct way to stop the problem with access tokens is by audience restricting them.
>> To do that the client needs to provide an audience in it’s request and the AS needs to include that in the AT.
>> I included the Authorization_endpoint URI in the hash to detect if the auth request may have been tampered with to change the audience for the AT.
>> For implicit we could have a version of PKCE where the AS returns a parameter with S256( client_id || authorization_endpoint URI || resource endpoint URI) to verify the request was not tampered with, and that would allow the AT to be properly audience restricted.
>> This would be a completely new approach not involving discovery, logical names for AS, or link relations.
>> Effectively this code_challenge method becomes a signature over parts of the request and the implicit audience of the token_endpoint URI.
>> People keep asking why PKCE doesn’t stop these attacks, and it won’t with the current PKCE methods,  however a new method along the lines I sketched out may let it work, or I could be completely wrong.
> Once you have written down something more definite I promise to implement it promptly such that we can do interop
> testing early in the process.
> Before John brought up this new PKCE mode I was on the Option A side of the fence now I’d like to see more
> of this PKCE idea before committing to one or the other.
> — Roland
> ”Everybody should be quiet near a little stream and listen."
> From ’Open House for Butterflies’ by Ruth Krauss