Re: [OAUTH-WG] New Version Notification for draft-campbell-oauth-tls-client-auth-00.txt

[Samuel Erdtman <[e-mail]>]

I can see your point, maybe the client_id will not be in the certificate.
If I had an AS I would select to trust one or several CAs and then create
certificate mappings between certificate serial number (or some other
unique attribute in the certificate) and client_id. If I were to bind a
specific certificate to a client_id I lose the flexibility of the PKI
(maybe what you want).

I think multiple certificates might not be a uncommon situation especially
if you call ASs from different organizations because they will trust
different CAs.

//Samuel


On Thu, Nov 3, 2016 at 5:32 PM, Justin Richer <[e-mail]> wrote:

> Jim,
>
> In those circumstances, are the clients generally calling multiple
> different services? Or just one? For those that call multiple services, are
> they using multiple (different) client certificates?
>
> I’m not saying the client would issue its own cert in all cases — much
> more common is what I’ve seen, with clients being assigned a certificate
> from a trusted CA, and then services that the client talks to being told to
> trust that CA and also assign the CN/DN of the cert a set of privileges.
> What I *haven’t* seen is a client being issued multiple certificates to
> talk to multiple systems. That latter case is common enough in the OAuth
> world that I wouldn’t want us to paint ourselves in a corner.
>
>  — Justin
>
> On Nov 3, 2016, at 10:31 AM, Jim Manico <[e-mail]> wrote:
>
> Thanks Justin. I use several security intel services and they all have
> different cert delivery mechanisms for mutual TLS. It's •rare• for services
> to let clients choose certs, they are usually assigned to users by each
> service from my experience.
>
> Aloha,
> --
> Jim Manico
> @Manicode
> Secure Coding Education
> [phone]
>
> On Nov 3, 2016, at 8:51 AM, Justin Richer <[e-mail]> wrote:
>
> Yes, I elided the certificate issuance process. The point remains the
> same: you're not going to be submitting a CSR to the same party you're
> getting your client_id from, usually. If the draft assumes that, then it's
> incredibly limiting.
>
>
> Do people really use separate TLS client certs for separate connections in
> the wild? I've personally never seen that. What I've seen is that a piece
> of software gets its certificate that it uses to make whatever connections
> it needs to make.
>
>
>  -- Justin
>
> On 11/3/2016 8:48 AM, Jim Manico wrote:
>
> Just to be clear, the relationship should more like...
>
> AS issues public key to clients, or client sends public key to AS. The
> authorities job is NOT to give the client the public key, but to sign the
> public key for authenticity. It's bad practice to accept the full cert (pub
> key+signature) from an authority. If an authority is creating your public
> key, they are also creating your private key.... bad.
>
> > The client will use the same cert across multiple connections, possibly
> multiple AS's, but the same isn't true of the client_id.
>
> This seems like a bad idea. I suggest a separate key/signature for each
> service.
> --
> Jim Manico
> @Manicode
> Secure Coding Education
> [phone]
>
> On Nov 3, 2016, at 8:41 AM, Justin Richer <[e-mail]> wrote:
>
> I agree that the client_id is unlikely to be found inside the certificate
> itself. The client_id is issued by the authorization server for the client
> to use at that single AS. The certificate is issued by the CA for the
> client to use on any connection. The AS and CA are not likely to be the
> same system in most deployments. The client will use the same cert across
> multiple connections, possibly multiple AS's, but the same isn't true of
> the client_id.
>
> Additionally, I think we want to allow for a binding of a self-signed
> certificate using dynamic registration, much the way that we already allow
> binding of a client-generated JWK today.
>
> I do think that more examples and guidance are warranted, though, to help
> AS developers.
>
>  -- Justin
>
> On 11/2/2016 5:03 PM, Brian Campbell wrote:
>
>
> On Sun, Oct 30, 2016 at 9:27 AM, Samuel Erdtman <[e-mail]> wrote:
>
>>
>> I agree it is written so that the connection to the certificate is
>> implicitly required but I think it would be better if it was explicit
>> written since the lack of a connection would result in a potential security
>> hole.
>>
>
> That's fair. I agree it can be made more explicit and that it be good to
> do so.
>
>
>
>> When it comes to the client_id I think subject common name or maybe
>> subject serial numbers will be the common location, and I think an example
>> would be valuable.
>>
>>
>
> In my experience and the way we built support for mutual TLS OAuth client
> auth the client_id value does not appear in the certificate anywhere. I'm
> not saying it can't happen but don't think it's particularly common.
>
> I can look at adding some examples, if there's some consensus that they'd
> be useful and this document moves forward.
>
>
>
>>
>> I´m not saying it is a bad Idea just that I would prefer if it was not a
>> MUST.
>> With very limited addition of code it is just as easy to get the
>> certificate attribute for client id as it is to get it from the HTTP
>> request data (at least in java). I also think that with the requirement to
>> match the incoming certificate in some way one has to read out the
>> certificate that was used to establish the connection to do some kind of
>> matching.
>>
>>
> Getting data out of the certificate isn't a concern. I just believe that
> the constancy of having the client id parameter is worth the potential
> small amount duplicate data in some cases. It's just a -00 draft though and
> if the WG wants to proceed with this document, we seek further input and
> work towards some consensus.
>
>
>
> _______________________________________________
> OAuth mailing [e-mail]://www.ietf.org/mailman/listinfo/oauth
>
>
> _______________________________________________
> OAuth mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/oauth
>
>
>
>