[keycloak-dev] JWS signatures using PS256 or ES256 algorithms for signing

[Marek Posolda]

On 28/06/18 10:12, Stian Thorgersen wrote:
>
>
> On Wed, 27 Jun 2018 at 09:31, Marek Posolda <mposolda at redhat.com 
> <mailto:mposolda at redhat.com>> wrote:
>
>     +1
>
>     Having the configuration in adapter (keycloak.json) won't scale
>     also because REST services (bearer-only clients) will need to
>     verify signatures by "frontend" clients, but each frontend client
>     can use different algorithm to sign it's token. So keycloak.json
>     of bearer-only clients would need to have some configuration map
>     to track all of this... Also "rotation" of algorithms will be
>     impossible (if administrator changes the algorithm for client in
>     admin console, the adapter configuration in keycloak.json will
>     need to be updated and hence application restarted...)
>
>     But I think we don't need to add any new endpoint, but just re-use
>     existing client registration endpoints for that? OIDC Client
>     registration has various metadata for clients like
>     id_token_signed_response_alg, id_token_encrypted_response_alg,
>     userinfo_signed_response_alg etc. See [1]. Also we have support
>     for Dynamic client registration management [2], so client
>     applications are able to "download" the client's metadata from the
>     endpoint.
>
>
> I wasn't thinking adding a new endpoint. We should use OIDC well-known 
> endpoint. However, that doesn't include any attributes for the access 
> token, only ID token. Which is of course due to the fact that the 
> access token is opaque in OIDC. We could for instance add 
> "access_token_signed_response_alg".
>
> Clients would by default check the OIDC well-known endpoint for what 
> algorithm to use, but it should also be possible to override in 
> keycloak.json as that OIDC well-known would say the realm default, but 
> we want to be able to override the algorithm for individual clients.
IMO OIDC well-known endpoint is not good for that. It's not the endpoint 
to be useful for retrieve metadata of specific client, rather it's 
global for a realm. Even for ID token, there is not single signature 
algorithm. There is just list of supported algorithms in 
id_token_signing_alg_values_supported .

I wonder that for our adapters, we can just use the Client Registration 
(Management) endpoint of our own representation - the endpoint of 
DefaultClientRegistrationProvider.getDefault() . That's our own rep, so 
we are fine to return the accessToken signature algorithm. We can ensure 
that algorithm is always returned (We will return client specific 
algorithm or fallback to realm default algorithm if client doesn't have 
one set). We can invoke that endpoint with the bearer-token, which 
adapter will always have - it can just use the token it wants to verify 
signature for. Any reason why it won't work?

I wouldn't add anything to keycloak.json. IMO it would suck because:
- If algorithm is changed by admin in KC admin console, adapter will 
need to be re-configured in keycloak.json and restarted (no automatic 
algorithm "rotation" support)
- Configuration of algorithm will be duplicated on both adapter and server
- Bearer-only clients can be invoked with tokens of various "frontend" 
clients signed by various different algorithms. So the configuration in 
keycloak.json would be quite complicated here.

Marek

>
>     The specification enforces that requests are authenticated by
>     Client Registration Access Token. But ATM we also support
>     authentication by bearer tokens, which is great. When application
>     sees the bearer token signed/encrypted by unknown algorithm, it
>     can send the request to KC registration/management endpoint of
>     particular client to download the client's metadata. This will
>     work fine also for bearer-only clients. The bearer-only client can
>     parse the token and send the request to the
>     registration/management endpoint of particular "frontend" client
>     the token was issued for to determine the right algorithms etc.
>
>
> Same as well-known. Dynamic client registration metadata from OIDC 
> doesn't specify anything for access token sign algorithm. So we would 
> have to add our own.
>
>
>     [1]
>     https://openid.net/specs/openid-connect-registration-1_0.html#ClientMetadata
>     [2] https://tools.ietf.org/html/draft-ietf-oauth-dyn-reg-management-05
>
>     Marek
>
>     On 26/06/18 18:25, Stian Thorgersen wrote:
>>     I don't really like the idea of having to reconfigure to make the
>>     adapter accept new signature. I know oidc well known  endpoint
>>     doesn't have signature algorithm for access token, but we could
>>     add one and have adapters pull from the server what algorithms to
>>     accept.
>>
>>     On Tue, 26 Jun 2018, 11:00 Marek Posolda, <mposolda at redhat.com
>>     <mailto:mposolda at redhat.com>> wrote:
>>
>>         On 30/05/18 09:35, Stian Thorgersen wrote:
>>>>         I think it might be better to determine which kind of Token Signature
>>>>         Provider be used by not parsing JWS, for example, looking up Client or
>>>>         Realm settings.
>>>>         This PR might have impacts on keycloak's performance because it has parsed
>>>>         JWS to determine it every time keycloak receives JWS Token.
>>>>
>>>         On the server-side that is easy. On the adapter side that would probably
>>>         require adding a property to keycloak.json to set the algorithm. In either
>>>         case it should probably default to RSA for existing realms at least, but we
>>>         could consider setting it to ES256 for new realms.
>>>
>>         +1
>>
>>         Parsing token signature to determine algorithm should be
>>         avoided IMO. AFAIR Some OAuth/OIDC vendors had security
>>         issues in the past, that they parsed the header with "none"
>>         algorithm and then client applications automatically trust
>>         unsigned tokens. We should make sure this is not possible.
>>
>>         Marek
>>
>
>