Re: [OAUTH-WG] [EXTERNAL] Re: dpop_jkt Authorization Request Parameter

[Warren Parad <[e-mail]>]

While I agree this is a problem, adding the thumbprint to the authorization
request only increases the difficulty for the malicious MITM attack to need
to also intercept the authorization request in addition to the token
request to swap out the dpop_jkt as well. If I'm right, then it doesn't
make sense to implement this as the solution.

While I agree, we could move the dpop_jwk determination to the
authorization request, instead of the token, it doesn't solve the problem.
What it does say is that the authorization request user-agent is the one
that has the key, and not the one doing the code exchange. Well this is
actually weird in the case of non-public clients, because it doesn't make
sense from that client perspective, as the "front-end" would need to now
have the constructed dpop_jkt even though it doesn't have the dpop key.

Warren Parad

Founder, CTO
Secure your user data with IAM authorization as a service. Implement
Authress <https://authress.io/>.


On Fri, Dec 3, 2021 at 12:22 AM Mike Jones <Michael.Jones=
[e-mail]> wrote:

> Thanks for this thoughtful analysis, Aaron.  I believe you’re spot on that
> these attacks can occur “when the attacker has access to both the
> authorization code as well as the PKCE code verifier.”
>
>
>
>                                                        -- Mike
>
>
>
> *From:* OAuth <[e-mail]> *On Behalf Of * Aaron Parecki
> *Sent:* Thursday, December 2, 2021 2:58 PM
> *To:* Warren Parad <wparad=[e-mail]>
> *Cc:* Pieter Kasselman <pieter.kasselman=[e-mail]>;
> [e-mail]
> *Subject:* Re: [OAUTH-WG] [EXTERNAL] Re: dpop_jkt Authorization Request
> Parameter
>
>
>
> Hi all, I've been giving this some more thought.
>
>
>
> The problem occurs when the attacker has access to both the authorization
> code as well as the PKCE code verifier. The assumption being made with PKCE
> is that the first time the PKCE code verifier and authorization code are
> seen together is in the POST request to the token endpoint. That means if
> there is a way to observe this request, the attacker can complete the
> exchange and get an access token.
>
>
>
> Pieter's writeup in the PDF describes one way that can happen, by
> exfiltrating log files. While I still agree with the sentiment in this
> thread that this is a relatively obscure condition, I also agree that it is
> actually something that can happen in the wild, especially since there are
> entire companies built around the idea of real-time log file analysis.
>
>
>
> That said, there are other ways an attacker could get access to these two
> pieces of information. What are the different points in a request lifecycle
> that could be attacked? The beginning, the middle and the end. We've talked
> about attacking the end, which is the log file example. Attacking the
> middle involves being in the middle of the TLS connection, which we also
> know is possible with corporate network proxies and such. We haven't yet
> talked about the beginning of the request. What can observe the beginning
> of a request? Here's a concrete example:
>
>
>
> Assume the OAuth client is a single-page app in a browser. The user is
> using an ad blocker installed as a browser extension. The ad blocker can be
> configured to observe and block network requests before they are made. If
> the extension is configured to attack a particular OAuth server, the JS
> client would make the token request containing the authorization code and
> PKCE code verifier, then the extension would be able to observe that
> request, block it, and ship the two values to the attacker's server where
> they can be redeemed and associated with the attacker's own DPoP key. Even
> perfectly single-use authorization codes don't help here either, because
> the original request was completely blocked.
>
>
>
> (Sidenote: To get ahead of any counterarguments here, yes, Chrome is
> eventually migrating to their new "manifest v3" which deprecates the
> webRequest API that allows this observation in favor of a different API
> that lets the browser block requests without making the actual request data
> available to the extension, which I am assuming they are doing in no small
> part because of the possibility of what I just described. It sounds like
> Mozilla is going to follow suit, but I haven't found concrete confirmation
> of that. That said, it will be a while before these changes are rolled out
> and support for the (dangerous) webRequest API is fully dropped, so this is
> likely going to continue to be a potential attack vector for a few years at
> least.)
>
>
>
> Because of the ease of deployment of a malicious browser extension, I do
> believe this is an important attack vector to consider and is worth
> solving. I am going to save any judgment on the particular dpop_jwk
> parameter proposal for a different thread, but I wanted to at least get on
> the same page about the fact that this is something worth solving first.
>
>
>
> Aaron
>
>
>
>
>
> On Thu, Dec 2, 2021 at 10:38 AM Warren Parad <wparad=
> [e-mail]> wrote:
>
> The only mention of sophistication is this logical fallacy:
>
>  If this leading security company had been penetrated, it almost certainly
> took an incredibly sophisticated attack.
>
>
>
> But it leaves out exactly what that was. And it doesn't give any insight
> into how this attack at MS would have been prevented despite the supply
> chain vulnerability based on the second point that Aaron made. If they are
> able to get the auth code, why aren't they able to get the DPoP signature?
> And then send both of these?
>
>
> Further in this case, it doesn't even matter if the attacker gets the
> access token if that access token is bound to the client, because it's
> worthless without the DPoP key. That's a much more secure solution than
> issuing non-bound Bearer tokens as a response to the bound authorization
> code. And if Bearer tokens are being used instead of bound tokens, then
> those could still end up in the logs, and be exfiltrated.
>
>
>
> In OAuth, the client already needs to authenticate with the AS, the spec
> is SHOULD, and options the client_secret already. Adding in the DPoP
> signature into the request is duplicating auth. If we don't like the client
> auth mechanisms to the AS, we should directly provide an auth RFC
> recommending better alternatives than sending a symmetric client_secret
> back to the AS.
>
>
>
> *Warren Parad*
>
> Founder, CTO
>
> Secure your user data with IAM authorization as a service. Implement
> Authress <https://authress.io/>.
>
>
>
>
>
> On Thu, Dec 2, 2021 at 4:42 PM Pieter Kasselman <pieter.kasselman=
> [e-mail]> wrote:
>
> Thanks for the comments and engagement Warren.
>
>
>
> The attacks we described and the ideas on mitigations are born out of
> attack vectors we are observing in the wild. They are not negligible. We
> are seeing a new class of very sophisticated attackers, and if you’re
> interested, this article provides good context on capability and
> sophistication of the attackers Brad Smith: Inside Microsoft during the
> SolarWinds hack (fastcompany.com)
> <https://www.fastcompany.com/90672384/microsoft-president-brad-smith-solarwinds-exclusive>.
> We are sharing this with the hope that the industry will benefit from our
> experiences and incorporate it into standards and products. Attacks that
> seemed impossibly complex are not only possible, but have become probable.
>
>
>
> The proposed changes for DPoP are not meant to replace the need for
> one-time use tokens (single use tokens are preferable and we should
> continue to expect them), but instead address the limitations around
> implementing one-time use, especially at scale. The 60s window you mention
> below is sufficiently long to be exploited by these sophisticated attackers.
>
>
>
> Cheers
>
>
>
> Pieter
>
>
>
> *From:* OAuth <[e-mail]> *On Behalf Of *Warren Parad
> *Sent:* Wednesday 1 December 2021 15:29
> *To:* Pieter Kasselman <pieter.kasselman=[e-mail]>
> *Cc:* Mike Jones <Michael.Jones=[e-mail]>;
> [e-mail]
> *Subject:* Re: [OAUTH-WG] [EXTERNAL] Re: dpop_jkt Authorization Request
> Parameter
>
>
>
> (e.g. one-time use in a certain timeframe etc).
>
>
>
> Sure but couldn't we just reduce the lifetime? Even if the token isn't one
> time use, surely the reuse time is trivially short which would prevent
> against exfiltration of the necessary security tokens to issue the attack?
>
>
>
> I feel like the simpler solution will always win, which in this case is
> one-time use tokens, then the problem is moot, right? So this only comes
> into play if you want to allow token reuse in a time window. The previously
> suggested max allowed time window from OAuth 2.1 was 60s for auth codes. So
> we are saying that the attack surface is still too large, for the .01% of
> implementations that have multi-use tokens, and the .01% of implementations
> that use the maximum 60s reuse, and then the subset of those that aren't
> correctly scrubing their logs, and then the subset of those that have a
> vulnerability which allows for exfiltration of both those logged tokens and
> the logged PKCE verifier?
>
>
>
> Why are we making this more complicated for a majority of cases, which:
>
>    - Only have single use tokens
>    - Or Only have a very short lifetime
>    - Or Are already correctly sanitizing their logs
>    - Or Have defense in depth for their deployments.
>
> If the implementation is so insecure that none of those are happening,
> wouldn't the implementation for this functionality also be suspect for an
> opportunity for attack?
>
>
>
> I feel like we are justifying here that multi-use tokens are wrong, but
> still want a solution to use them. Once we've proven that an deployment is
> not okay with using multi-use tokens, then the conclusion should be "don't
> have multi-use tokens", not: "let's still have multi-use tokens, but come
> up with a complex way to prevent their multi-use from accidentally being
> abused".
>
>
>
> Or am I missing something that would actually make this a
> non-negligible attack vector?
>
>
>
> - Warren
>
>
>
> *Warren Parad*
>
> Founder, CTO
>
> Secure your user data with IAM authorization as a service. Implement
> Authress
> <https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fauthress.io%2F&data=04%7C01%7Cpieter.kasselman%40microsoft.com%7Cb5c71bfcbfbb48fd641508d9b4df5fcf%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739693847580905%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=%2FBkvuWZ3FVTcdTtfe%2FoLurIGxcsJHCz6zXmW1PROTSc%3D&reserved=0>
> .
>
>
>
>
>
> On Wed, Dec 1, 2021 at 4:14 PM Pieter Kasselman <pieter.kasselman=
> [e-mail]> wrote:
>
> Hi Aaron, Neil
>
>
>
> Thanks for the questions.
>
>
>
> We agree that ideally authorization codes and PKCE proofs would never end
> up in log files and one-time use would be perfectly implemented.
>
>
>
> However, in practice these artefacts do find their way into log files in
> various places and one-time use may not always be practical (e.g. one-time
> use in a certain timeframe etc).
>
>
>
> The addition of these mitigations is not meant to replace the need for
> one-time use or good logging hygiene. Instead they provide pragmatic
> defence in depth against real attacks rather than assuming perfect
> implementations. We are deploying these mitigations and are sharing them
> for inclusion in DPoP to enable others to do the same.
>
>
>
> Regarding the question about interrupting/intercepting the HTTPS
> connection, the attacker don’t need to intercept the HTTPS connection or
> modify the content in the TLS tunnel, rather they just need to prevent the
> authorization code from being presented to the Authorization Server. It may
> even happen due to a poor network connection. The poor connection may be
> engineered by an attacker, or they may opportunistically benefit from it.
> The networks are not perfect either.
>
>
>
> Cheers
>
>
>
> Pieter
>
>
>
>
>
> *From:* OAuth <[e-mail]> *On Behalf Of *Aaron Parecki
> *Sent:* Wednesday 1 December 2021 00:05
> *To:* Neil Madden <[e-mail]>
> *Cc:* Mike Jones <Michael.Jones=[e-mail]>;
> [e-mail]
> *Subject:* [EXTERNAL] Re: [OAUTH-WG] dpop_jkt Authorization Request
> Parameter
>
>
>
> I tend to agree with Neil here. I'm struggling to see the relevance of
> this attack.
>
>
>
> It seems like the PDF writeup describes two possible reasons an attacker
> could get access to the authorization code and PKCE code verifier.
>
>
>
> 1. The attacker has access to the logs of the token endpoint.
>
> 2. The attacker can intercept HTTPS connections between the client and AS
> (VPN, corporate network proxy, etc)
>
>
>
> For 1, the solution is to stop logging the contents of the POST body, and
> secure your infrastructure. I don't think making the client jump through
> extra hoops is a good solution if you are already logging more than you
> should be or you don't trust the people who have access to the
> infrastructure. If this really is a concern, I suspect there are a lot more
> places in the flow that would need to be patched up if you don't trust your
> own token endpoint.
>
>
>
> For 2, if the attacker can intercept the HTTPS connection, then the
> proposed solution doesn't add anything because the attacker could modify
> the requests before it hits the authorization server anyway, and change
> which DPoP key the token gets bound to in the first place. Plus, the
> attacker would also have access to anything else the client is sending to
> the AS, such as the user's password when they authenticate at the AS.
>
>
>
> Are there other attack vectors I'm missing that might actually be solved
> by this mechanism?
>
>
>
> Aaron
>
>
>
>
>
> On Tue, Nov 30, 2021 at 12:40 PM Neil Madden <[e-mail]>
> wrote:
>
> Sadly I couldn’t make the DPoP session, but I’m not convinced the attack
> described in the earlier message really needs to be prevented at all. The
> attack largely hinges on auth codes not being one-time use, which is not a
> good idea, or otherwise on poor network security on the token endpoint. I’m
> not convinced DPoP needs to protect against these things. Is there more to
> this?
>
>
>
> The proposed solutions also seem susceptible to the same problems they
> attempt to solve - if an attacker is somehow able to interrupt the client’s
> (TLS-protected) token request, why are they somehow not able to
> interrupt/modify the (far less protected) redirect to the authorization
> endpoint?
>
>
>
> — Neil
>
>
>
> On 30 Nov 2021, at 20:15, Mike Jones <
> Michael.Jones=[e-mail]> wrote:
>
>
>
> As described during the OAuth Security Workshop session on DPoP, I created
> a pull request adding the dpop_jkt authorization request parameter to use
> for binding the authorization code to the client’s DPoP key.  See
> https://github.com/danielfett/draft-dpop/pull/89
> <https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fdanielfett%2Fdraft-dpop%2Fpull%2F89&data=04%7C01%7Cpieter.kasselman%40microsoft.com%7Cb5c71bfcbfbb48fd641508d9b4df5fcf%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739693847580905%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=ASCRFFPMA7qIItkxpVTrVaJtC53R2niWOzB0l0GQKrw%3D&reserved=0>
> .
>
>
>
> This is an alternative to https://github.com/danielfett/draft-dpop/pull/86
> <https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fdanielfett%2Fdraft-dpop%2Fpull%2F86&data=04%7C01%7Cpieter.kasselman%40microsoft.com%7Cb5c71bfcbfbb48fd641508d9b4df5fcf%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739693847580905%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=lgb9WmnOwWVtNIFsZ1mQG4jSBQYLZv%2BETe6HIKFeerg%3D&reserved=0>,
> which achieved this binding using a new DPoP PKCE method.  Using this
> alternative allows PKCE implementations to be unmodified, while adding DPoP
> in new code, which may be an advantage in some deployments.
>
>
>
> Please review and comment.  Note that I plan to add more of the attack
> description written by Pieter Kasselman to the security considerations in a
> future commit.  This attack description was sent by Pieter yesterday in a
> message with the subject “Authorization Code Log File Attack (was DPoP
> Interim Meeting Minutes)”.
>
>
>
>                                                        -- Mike
>
>
>
> _______________________________________________
> OAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/oauth
> <https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ietf.org%2Fmailman%2Flistinfo%2Foauth&data=04%7C01%7Cpieter.kasselman%40microsoft.com%7Cb5c71bfcbfbb48fd641508d9b4df5fcf%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739693847580905%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=vuBY0pdcSiMXQF213ZVLm4yNMFhRqM1jWlrWSzn%2FS%2FE%3D&reserved=0>
>
>
>
> _______________________________________________
> OAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/oauth
> <https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ietf.org%2Fmailman%2Flistinfo%2Foauth&data=04%7C01%7Cpieter.kasselman%40microsoft.com%7Cb5c71bfcbfbb48fd641508d9b4df5fcf%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739693847580905%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=vuBY0pdcSiMXQF213ZVLm4yNMFhRqM1jWlrWSzn%2FS%2FE%3D&reserved=0>
>
> _______________________________________________
> OAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/oauth
> <https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ietf.org%2Fmailman%2Flistinfo%2Foauth&data=04%7C01%7Cpieter.kasselman%40microsoft.com%7Cb5c71bfcbfbb48fd641508d9b4df5fcf%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739693847580905%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=vuBY0pdcSiMXQF213ZVLm4yNMFhRqM1jWlrWSzn%2FS%2FE%3D&reserved=0>
>
> _______________________________________________
> OAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/oauth
>
> _______________________________________________
> OAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/oauth
>