ContainerImage.Pinniped/proposals/1125_dynamic-supervisor-oidc-clients

title	authors	status	sponsor	approval_date
Dynamic Supervisor OIDC Clients	@cfryanr	in-review

Disclaimer: Proposals are point-in-time designs and decisions. Once approved and implemented, they become historical documents. If you are reading an old proposal, please be aware that the features described herein might have continued to evolve since.

Dynamic Supervisor OIDC Clients

Problem Statement

Pinniped can be used to provide authentication to Kubernetes clusters via kubectl for cluster users such as developers, devops teams, and cluster admins. However, sometimes these same users need to be able to authenticate to webapps running on these clusters to perform actions such as installing, configuring, and monitoring applications on the cluster. It would be fitting for Pinniped to also provide authentication for these types of webapps, to ensure that the same users can authenticate in exactly the same way, using the same identity provider, and resolving their identities to the same usernames and group memberships. Enabling this use case will require new features in Pinniped, which are proposed in this document.

How Pinniped Works Today (as of version v0.15.0)

Each FederationDomain configured in the Pinniped Supervisor is an OIDC Provider issuer which implements the OIDC authorization code flow.

Today, the Pinniped Supervisor only allows one hardcoded OIDC client, called pinniped-cli. This client is only allowed to redirect authcodes to the CLI's localhost listener. This makes it intentionally impossible for this client to be used by a webapp running on the cluster (or anywhere else on the network). The pinniped-cli client is implicitly available on all FederationDomains configured in the Supervisor, since every FederationDomain allows users to authenticate themselves via the Pinniped CLI for kubectl integration.

Terminology / Concepts

• See the definition of a "client" in the OAuth 2.0 spec. For the purposes of this proposal, a "client" is roughly equal to a webapp which wants to know the authenticated identity of a user, and may want to perform actions as that user on clusters. An admin needs to allow the client to learn about the identity of the users by registering the client with the Pinniped Supervisor.
• See also the OIDC terminology in the OIDC spec.
• The OIDC clients proposed in this document are "dynamic" in the sense that they can be configured and reconfigured on a running Supervisor by the admin.

Proposal

Goals and Non-goals

Goals for this proposal:

• Allow Pinniped admins to configure applications (OIDC clients) other than the Pinniped CLI to interact with the Supervisor.
• Provide a mechanism which governs a client's access to the token exchange APIs. Not all webapps should have permission to act on behalf of the user with the Kubernetes API of the clusters, so an admin should be able to configure which clients have this permission.
• Provide a mechanism for requesting access to different aspects of a user identity, especially getting group memberships or not, to allow the admin to exclude this potentially information for clients which do not need it.
• Support a web UI based LDAP/ActiveDirectory login screen. This is needed to avoid having webapps handle the user's password, which should only be seen by the Supervisor and the LDAP server. However, the details of this item have been split out to a separate proposal document.
• Client secrets should be stored encrypted or hashed, not in plain text.

Non-goals for this proposal:

• Pinniped's scope is to provide authentication for cluster users. Providing authentication for arbitrary users to arbitrary webapps is out of scope. The only proposed use case is providing the exact same identities that are provided by using Pinniped's kubectl integration, which are the developers/devops/admin users of the cluster.
• Supporting any OAuth/OIDC flow other than OIDC authorization code flow.
• Implementing any self-service client registration API. Clients will be registered by the Pinniped admin user.
• Implementing a consent screen. This would be clearly valuable but will be left as a potential future enhancement in the interest of keeping the first draft of this feature smaller.
• Management (ie creation & rotation) of client credentials on the operator's behalf. This will be the operator's responsibility.
• Orchestration of token exchanges on behalf of the client. Webapps which want to make calls to the Kubernetes API of clusters acting as the authenticated user will need to perform the rest of the token and credential exchange flow that it currently implemented by the Pinniped CLI. Providing some kind of component or library to assist webapp developers with these steps might be valuable but will be left as a potential future enhancement.

Specification / How it Solves the Use Cases

This document proposes supporting a new Custom Resource Definition (CRD) for the Pinniped Supervisor which allows the admin to create, update, and delete OIDC clients for the Supervisor.

API Changes

Configuring clients

An example of the new CRD to define a client:

apiVersion: oauth.supervisor.pinniped.dev/v1alpha1
kind: OIDCClient
metadata:
  name: my-webapp-client
  namespace: pinniped-supervisor
spec:
  id: my-webapp
  secretNames:
    - my-webapp-client-secret-1
    - my-webapp-client-secret-2
  allowedRedirectURIs:
    - https://my-webapp.example.com/callback
  allowedGrantTypes:
    - authorization_code
    - refresh_token
    - urn:ietf:params:oauth:grant-type:token-exchange
  allowedScopes:
    - openid
    - offline_access
    - pinniped:request-audience
    - groups
status:
  conditions:
    - type: ClientIDValid
      status: False
      reason: InvalidCharacter
      message: client IDs are not allowed to contain ':'

A brief description of each field:

• name: Any name that is allowed by Kubernetes.
• namespace: Only clients in the same namespace as the Supervisor will be honored. This prevents cluster users who have write permission in other namespaces from changing the configuration of the Supervisor.
• id: The client ID, which is conceptually the username of the client. Validated against the same rules applied to name. Especially note that : characters are not allowed because the basic auth specification disallows them in usernames.
• secretNames: The names of Secrets in the same namespace which contain the client secrets for this client. A client secret is conceptually the password for this client. Clients can have multiple passwords at the same time, which are all acceptable for use during an authcode flow. This allows for smooth rotation of the client secret by an admin without causing downtime for the webapp's authentication flow.
• allowedRedirectURIs: The list of allowed redirect URI. Must be https:// URIs, unless the host of the URI is 127.0.0.1, in which case http:// is also allowed (see RFC 8252).
• allowedGrantTypes: May only contain the following valid options:
  • authorization_code allows the client to perform the authorization code grant flow, i.e. allows the webapp to authenticate users.
  • refresh_token allows the client to perform refresh grants for the user to extend the user's session.
  • urn:ietf:params:oauth:grant-type:token-exchange allows the client to perform RFC8693 token exchange, which is a step in the process to be able to get a cluster credential for the user.
• allowedScopes: Decide what the client is allowed to request. Note that the client must also actually request particular scopes during the authorization flow for the scopes to be granted. May only contain the following valid options:
  • openid: The client is allowed to request ID tokens.
  • offline_access: The client is allowed to request an initial refresh token during the authorization code grant flow.
  • pinniped:request-audience: The client is allowed to request a new audience value during a RFC8693 token exchange, which is a step in the process to be able to get a cluster credential for the user.
  • groups: The client is allowed to request that ID tokens contain the user's group membership, if their group membership is discoverable by the Supervisor. This is a newly proposed scope which does not currently exist in the Supervisor. Without the groups scope being requested and allowed, the ID token would not contain groups.
• conditions: The result of validations performed by a controller on these CRs will be written by the controller on the status.

Some other settings are implied and will not be configurable:

• All clients must use PKCE during the authorization code flow. There is a risk that some client libraries might not support PKCE, but it is considered a modern best practice for OIDC.
• All clients must use client secret basic auth for authentication at the token endpoint. This is the most widely supported authentication method in client libraries and is recommended by the OAuth 2.0 spec over the alternative of using query parameters in a POST body.
• All clients are only allowed to use code as the response_type at the authorization endpoint.
• All clients are only allowed to use the default or specify query as the response_mode at the authorization endpoint. This excludes clients from using form_post. We could consider allowing form_post in the future if it is desired.
• Clients are not allowed to use JWT-based client auth. This could potentially be added as a feature in the future.

Configuring clients via an aggregated API server (alternative data model to CRD)

CRDs are convenient to use, however, they have one core limitation - they cannot represent non-CRUD semantics. For example, the existing token credential request API could not be implemented using a CRD - it processes an incoming token and returns a client certificate without writing any data to etcd. Aggregated APIs have no such limitation.

An example of how the aggregated API would define a client:

apiVersion: oauth.supervisor.pinniped.dev/v1alpha1
kind: OIDCClient
metadata:
  name: my-webapp-client
  namespace: pinniped-supervisor
spec:
  allowedRedirectURIs:
    - https://my-webapp.example.com/callback
  allowedGrantTypes:
    - authorization_code
    - refresh_token
    - urn:ietf:params:oauth:grant-type:token-exchange
  allowedScopes:
    - openid
    - offline_access
    - pinniped:request-audience
    - groups

Some key differences from the CRD based approach:

• The .status field is omitted - validation is performed statically on every create and update operation - the client object is always guaranteed to be valid. The admin is notified directly via a validation error if they attempt to create an incorrect object (for example, with metadata.name set to a value that includes a :). Similarly, if the .spec.allowed* fields include invalid data, the object will be rejected. Consumers of this API, such as our controllers, can rely on the data being valid, which simplifies their logic.
• .spec.secretNames has been removed as the client secret is generated on the server after the client has already been created. This is handled via a create operation to the secret subresource (this is modeled after the Kubernetes service account API's token subresource). Just as the SA token subresource uses the TokenRequest API as the input and output schema, the secret subresource would use a SecretRequest API as the input and output schema. The response from this API is the only time that the plaintext client secret is made available. Since rotation is not part of this proposal, the secret API can only be invoked once per client - all subsequent requests will fail. A "hard" rotation can be performed by deleting and re-creating the client, followed by calling the secret API. Note that this purely an artificial limitation - it would not be difficult to allow multiple invocations of the secret API (with each creating a new client secret). SecretRequestSpec would need to be expanded with a bool field to revoke old client secrets (all except the latest, a new client secret would not be generated for this call).

type SecretRequest struct {
  metav1.TypeMeta   `json:",inline"`
  metav1.ObjectMeta `json:"metadata,omitempty"`


  Spec SecretRequestSpec     `json:"spec"`
  Status SecretRequestStatus `json:"status"`
}

type SecretRequestSpec struct {}  // empty for now

type SecretRequestStatus struct {
  Secret string `json:"secret"`
}

The bulk of the aggregated API server implementation would involve copy-pasting the concierge aggregated API server code and changing the API group strings. The interesting part of the implementation would be the rest storage implementation of the OIDCClient API. Normally this would be done via direct calls to etcd, but running etcd is onerous. Instead we would follow the same model we use for the fosite storage layer - everything would be stored in Kubernetes secrets. The Kubernetes API semantics such as resource version would be trivial to implement since we would simply passthrough the object meta information from the underlying secret (each client would have a 1:1 mapping with a secret). The existing crud.Storage code can likely be re-used for this purpose with minor modifications. Each secret would store a single JSON object with the following schema:

type clientStorage struct {
  Name string              `json:"name"`
  Labels map[string]string `json:"labels"`
  Client OIDCClientSpec    `json:"client"`
  SecretHashes []string    `json:"hashes"`
  Version string           `json:"version"`  // updating this would require some form of migration
}

The client field would store the .spec provided by the user. The hashes field would store the hash of the server generated client secret on calls to the secret subresource API. The name field would store the metadata.name of the OIDCClient as the actual secret would be named pinniped-storage-oidcclient-<base32(sha256Hash(metadata.name))> to prevent any issues that could arise from giving the user control over the name (i.e. length issues, collision issues, validation issues, etc). Similarly, the labels field would store the metadata.labels of the OIDCClient to prevent any collisions with our internal labels such as storage.pinniped.dev/type. While the bulk of the object meta's fields such as resourceVersion and creationTimestamp would be a simple passthrough, others would not be supported such as generateName and managedFields (there is limited value in doing so, at least for the MVP). All of the standard Kubernetes rest verbs would be implemented.

Configuring client secrets

We wish to avoid storage of client secrets (passwords) in plain text. They should be stored encrypted or hashed.

Perhaps the most common approach for this is to use bcrypt with a random salt and a sufficiently high input cost. The salt protects against rainbow tables, and the input cost provides some protection against brute force guessing when the hashed password is leaked or stolen. However, the input cost also makes it slower for users to authenticate. The cost should be balanced against the current compute power available to attackers versus the inconvenience to users caused by a long pause during a genuine login attempt. There is no "best" value for the input cost. Even when an administrator determines a value that works for them, they should reevaluate as Moore's Law (and the availability of specialized hardware) catches up to their choice later.

Client secrets should be decided by admins. Many OIDC Providers auto-generate client secrets and return the generated secret once (and only once) in their API or UI. This is good for ensuring that the secret contains a large amount of entropy by auto-generating long random strings using lots of possible characters. However, Pinniped has declarative configuration as a design goal. The configuration of Pinniped should be able to be known a priori (even before installing Pinniped) and should be easy to include in a Gitops workflow.

Even if the client secrets are hashed with bcrypt, the hashed value is still very confidential, due to the opportunities for brute forcing provided by knowledge of the hashed value. Confidential data in Kubernetes should be stored in Secret resources. This makes it explicit that the data is confidential and many Kubernetes workflows are built on this assumption. For example, deployment tools will avoid showing the values in Secrets during an application deployment. As another example, Kubernetes best practices suggest that admins should use authorization policies to restrict read permission to Secrets as much as possible. Additionally, some clusters may use the Kubernetes feature to encrypt Secrets at rest, and thus reasonably expect that all confidential data is encrypted at rest.

Option 1: Providing client secrets already hashed

An admin could run bcrypt themselves to hash their desired client secret. Then they could write the resulting value into a Secret.

apiVersion: v1
kind: Secret
metadata:
  name: my-webapp-client-secret-1
  namespace: pinniped-supervisor
type: secrets.pinniped.dev/oidc-client-secret-bcrypt
stringData:
  clientSecret: $2y$10$20UQo9UzqBzT.mgDRu9TwOC...EQSbS2

Advantages:

• Least implementation effort.
• Admins choose their own preferred bcrypt input cost.
• Confidential data is stored in a Secret.

Disadvantages:

• Running bcrypt is an extra step for admins or admin process automation scripts. However, there are many CLI tools available for running bcrypt, and every popular programming language has library support for bcrypt. For example, htpasswd is pre-installed on all MacOS machines and many linux machines, and tools like Bitnami's bcrypt-cli are readily available. E.g. the following command generates and hashes a strong random password using an input cost of 12: p="$(openssl rand -hex 14)" && echo "$p" && echo -n "$p" | bcrypt-cli -c 12.

Option 2: Providing client secrets in plaintext and then automatically hashing them

An admin could provide the plaintext client secrets on the OIDCClient CR, instead of listing references to Secrets. A dynamic mutating admission webhook could automatically run bcrypt on each incoming plaintext secret, store the results somewhere, and remove the plaintext passwords.

There are several places that the hashed passwords could be stored by the webhook:

• In the same field on the OIDCClient CR, by replacing the plaintext passwords with hashed passwords
• In a different field on the OIDCClient CR, and deleting the plaintext passwords
• In Secret resources, by deleting the plaintext passwords and adding secretRefs to the OIDCClient CR, and creating/updating/deleting Secrets named with random suffixes

Advantages:

• The admin does not need to run bcrypt themselves.

Disadvantages:

• The development cost would be higher.
  • Pinniped does not currently have any admission webhooks, and they are not the simplest API to implement correctly.
  • Webhooks should use TLS, so Pinniped would need code to automatically provision a CA and TLS certs, and a controller to update the webhook configuration to have the generated CA bundle.
  • The webhook should also use mTLS (or a bearer token) to authenticate that requests are coming from the API server, which is another additional amount of effort similar to TLS.
  • The webhook should not be available outside the cluster, so it should be on a new listening port with a new Service.
• If the webhook goes down or has a bug, then all edits to the CR will fail while the issue is happening.
• Confidential data should be stored in a Secret, making the options to store the hashed passwords on the OIDCClient CR less desirable. Having an admission controller for Secrets would be putting Pinniped into the critical path for all create/update operations of Secrets in the namespace, which is probably not desirable either, since the admin user already creates lots of other Secrets in the namespace, and the Supervisor itself also creates many Secrets (e.g. for user session storage). This only leaves the option of having the webhook create side effects by watching OIDCClient CRs but mutating Secrets. The Kubernetes docs explain several complications that must be handled by webhooks that cause side effects.
• The desired semantics of this webhook's edit operations are not totally clear.
  • If a user updates the passwords or updates some unrelated field, then how would the webhook know to avoid regenerating the hashed passwords for unchanged passswords while updating/deleting the hashed passwords for those that changed or were deleted? Passwords are hashed with a random salt, so the incoming plaintext password would need to be compared against the hash using the same operation as when a user is attempting a login, which is purposefully slow to defeat brute-force attacks. Webhooks must finish within 10-30 seconds (10 seconds is the default timeout, and 30 seconds is the maximum configuration value for timeouts). In the event that the webhook determines that it should hash the passwords to store them, that is another intentionally slow operation. One can imagine that if there are three passwords and each takes 2 seconds to hash to determine which need to change, and then those that need to be updated take another 2 seconds to actually update, then the 10-second limit could be easily exceeded.
  • If a user reads the value of the CR (which never returns plaintext passwords) and writes back that value, does that delete all the hashed passwords? It would appear to the webhook that the admin wants to remove all client secrets.
  • Note that the Kubernetes docs say, "Mutating webhooks must be idempotent, able to successfully process an object they have already admitted and potentially modified." So the webhook would need to somehow recognize that it does not need to perform any update after it has already removed the plaintext passwords.
• Pinniped would need to offer an additional configuration option for the bcrypt input cost. There is no "correct" value to use unless it is determined by the admin user.
• The Kubernetes docs say, "It is recommended that admission webhooks should evaluate as quickly as possible, typically in milliseconds, since they add to API request latency. It is encouraged to use a small timeout for webhooks." Evaluating in milliseconds will not be possible due to the intentional slowness of bcrypt.
• The Kubernetes docs warn that current and future control loops may break when changing the value of fields. The docs say, "Built in control loops may break when the objects they try to create are different when read back. Setting originally unset fields is less likely to cause problems than overwriting fields set in the original request. Avoid doing the latter." So removing or updating an incoming plaintext password field is not advised, although the advice is not very specific.

Option 3: Aggregated API

See the earlier section where an aggregated API is discussed. Aggregated APIs solve both the data model and the secret provisioning problem.

Option 4: Custom REST endpoint for secret provisioning

This approach combines CRD based storage with the aggregated API subresource secret generation approach. While CRDs do not support subresources today, this can be emulated by either creating an aggregated API or a custom rest endpoint. The aggregated API approach would be similar to the subresource approach described above, notably with the API server handling authentication. The easier approach would be to implement the SecretRequest API as an endpoint on the supervisor such as https://<supervisor_url>/secret_request_v1alpha1/<client_name>. Authentication would be handled in a similar way to the token exchange API - a pinniped access token would be used to validate the identity of the caller. Unlike the token exchange API, this endpoint would require authorization. This would be handled by issuing a subject access review for the access token's identity (note that the virtual verb request is used to distinguish this API from regular Kubernetes rest APIs):

{
  "name": "my-webapp-client",
  "namespace": "pinniped-supervisor",
  "verb": "request",
  "group": "oauth.supervisor.pinniped.dev",
  "version": "*",
  "resource": "oidcclients",
  "subresource": "secretrequest"
}

Calls to the secret_request_v1alpha1 endpoint would, after successful auth, fetch the client_name OIDC client from the Kubernetes API (to validate that it exists and learn its UID), and then create a secret with an owner reference back to the OIDC client. This secret would be named pinniped-storage-oidcclientsecret-<base32(sha256Hash(client_name))> and would store the hash of the server generated secret. The details regarding rotation as mentioned for the aggregated subresource API implementation apply here as well. There would always be at most one Kubernetes secret per OIDC client. Since the secret name is deterministic based on the client name, no reference field is required on the OIDC client.

Option 5: Asymmetric crypto to store secret in status

TODO

Configuring association between clients and issuers

Each FederationDomain is a separate OIDC issuer. OIDC clients typically exist in a single OIDC issuer. Each OIDC client is effectively granted permission to learn about the users of that FederationDomain and to perform actions on behalf of the users of that FederationDomain. If the client is allowed by its configuration, then it may also perform actions against the Kubernetes APIs of all clusters associated with the FederationDomain.

It seems desirable for an admin to explicitly choose which clients are associated with a FederationDomain. For example, if an admin has a FederationDomain for all the Kubernetes clusters used by the finance department, and another FederationDomain for all the clusters used by the HR department, then a webapp for the finance department developers should not necessarily be allowed to perform actions on the Kubernetes API of the HR department's clusters.

Option 1: Explicitly associate specific clients with issuers

Each FederationDomain could list which clients are allowed. For example:

apiVersion: config.supervisor.pinniped.dev/v1alpha1
kind: FederationDomain
metadata:
  namespace: pinniped-supervisor
  name: my-domain
spec:
  issuer: https://my-issuer.pinniped.dev/issuer
  # This is the new part...
  clientRefs:
    - "my-webapp-client"
    - "my-other-webapp-client"

The pinniped-cli client does not need to be listed, since it only makes sense to allow kubectl access to all users of all FederationDomains. Additionally, the pinniped-cli can only redirect authcodes to localhost listeners, effectively only allowing users to log into their own accounts on their own computers.

Option 2: Implicitly associate all clients with all issuers

Rather than explicitly listing which clients are allowed on a FederationDomain, all FederationDomains could assume that all clients are available for use.

Advantages:

• Slightly less configuration for the user.
• Slightly less implementation effort since the FederationDomain watching controller would not need to change to read the list of clientRefs.

Disadvantages:

• Reusing the example scenario from above (finance and HR clusters), there would be no way to prevent a webapp for finance developer users from performing operations against the clusters of HR developer users who log into the finance webapp. This could be a serious security problem after the planned multiple identity providers feature is implemented to allow for more differentiation between the users of the two FederationDomains. This problem is compounded by the fact that many upstream OIDC providers use browser cookies to avoiding asking an active user to interactively log in again, and also by the fact that we decided to punt implementing a user consent UI screen. Together, these imply that an attacker from the finance department cluster which runs the client webapp would only need to trick an HR user into clicking on a single link in their web browser or email client for the attacker to be able to gain access to the HR clusters using the identity of the HR user, with no further interaction required by the HR user beyond just clicking on the link.

Upgrades

The proposed CRD will be new, so there aren't any upgrade concerns for it. Potential changes to the FederationDomain resource are also new additions to an existing resource, so there are again no upgrade concerns there.

The pinniped-cli client needs to continue to act as before for backwards compatibility with existing installations of the Pinniped CLI on user's machines. Therefore, it should be excluded from any new scope-based requirements which would restrict the group memberships from being returned. This will allow mixing of old CLIs with new Supervisors, and new CLIs with old Supervisors, in regard to the new Supervisor features proposed herein.

Tests

As usual, unit tests should be added for all new/changed code.

Integration tests should be added to mimic the usage patterns of a webapp. Dynamic clients should be configured with various options to ensure that the options work as expected. Those clients should be used to perform the authcode flow, RFC8693 token exchanges, and TokenCredentialRequest calls. Negative tests should include validation failures on the new CRs, and failures to perform actions that are supposed to be disallowed on the client by its configuration.

New Dependencies

None are foreseen.

Performance Considerations

Some considerations were mentioned previously for client secret option 2 above. No other performance impact is foreseen.

Observability Considerations

None are foreseen, aside from the usual error messages and log statements for the new features similar to what is already implemented in the Supervisor.

Security Considerations

Some security considerations were already mentioned above. Here are a couple more.

Ensuring reasonable client secrets

During a login flow, when the client secret is presented in plaintext to the Supervisor’s token endpoint, it could potentially validate that the secret meets some minimum entropy requirements. For example, it could check that the secret has sufficient length, a sufficient number of unique characters, and a sufficient number of letter vs number characters. If we choose to use the plaintext passwords option then the Supervisor could potentially perform this validation in the mutating admission webhook before it hashes the passwords.

Disallowing audience confusion

Who decides the names of the dynamic client and the workload clusters?

• The name of the dynamic client would be chosen by the admin of the Supervisor. We could put validations on the name if we would like to limit the allowed names.
• The name of the workload cluster is chosen by the admin of the workload cluster (potentially a different person or automated process). We don’t currently limit the string which chooses the audience name for the workload cluster, so it can be any non-empty string. The Supervisor is not aware of these strings in advance.

Given that, there are two kinds of potential audience confusion for the token exchange.

1. The ID token issued during the original authorization flow (before token exchange) will have an audience set to the name of the dynamic client. If this client’s name happened to be the same name as the name of a workload cluster, then the client could potentially skip the token exchange and use the original ID token to access the workload cluster via the Concierge (acting as the user who logged in). These initial ID tokens were not meant to grant access to any particular cluster.

   If the admin always names the dynamic clients in a consistent way which will not collide with the names of any workload cluster that they also name, then this won't happen unless another admin of a workload cluster breaks the naming convention. In that case, the admin of the workload cluster has invited this kind of token misuse on their cluster, possibly by accident. It is unlikely that it would be by accident if the naming convention of clusters included any random element, which is what the Pinniped docs recommend. This could either be solved with documentation advising against these naming collisions, or by adding code to make it impossible.

   We could consider inventing a way to make this initial ID token more inert. One possibility would be to take advantage of the RequiredClaims field of the JWT authenticator. The token exchange could be enhanced to always add a new custom claim to the JWTs that it returns, such as pinniped_allow_concierge_tcr: true. The Concierge JWTAuthenticator could be enhanced to require this claim by default, along with a configuration option to disable that requirement for users who are not using the Supervisor. ID tokens returned during an initial login (authcode flow) would not include this claim, rendering them unusable at the Concierge's TokenCredentialRequest endpoint. Docs could be updated to explain that users who configure dynamic clients should upgrade to use the version of the Concierge which performs this new validation on workload clusters, and that users who are using JWTAuthenticators for providers other than the Supervisor would need to add config to disable the new validation when they upgrade.

2. A user can use the public pinniped-cli client to log in and then to perform a token exchange to any audience value. They could even ask for an audience which matches the name of an existing dynamic client to get back an ID token that would appear as if it were issued to that dynamic client. Then they could try to find a way to use that ID token with a webapp which uses that dynamic client to authenticate its users (although that may not be possible depending on the implementation of the webapp). This could be prevented by making this an error in the token exchange code. No token exchange should be allowed if it requests that the new audience name be the name of any existing client in that FederationDomain, to avoid this kind of purposeful audience confusion.

Option 1: Add automatic prefix to the audience of ID tokens retrieved from token exchange

The audience parameter passed into the token exchange API is directly used as the audience for the resulting ID token. This could instead be changed to automatically attach the prefix client.oauth.pinniped.dev/token-exchange: to the requested audience, i.e. if audience=cluster-001 is passed, the resulting ID token will have an audience of client.oauth.pinniped.dev/token-exchange:cluster-001. This is guaranteed never to conflict with the audience of an OAuth client because clients cannot include : in their name (since that is not allowed in basic auth). Without further action, this is guaranteed to be a breaking change as existing JWTAuthenticator configurations will not work. Note in particular that some environments use old versions of the concierge with newer versions of the supervisor. This is not an officially supported or tested config, but we may want to consider this scenario.

An approach that can help mitigate backwards compatibility issues:

When the access token passed via the subject_token parameter was issued for the pinniped-cli OAuth client, the resulting audience will be set to both cluster-001 and client.oauth.pinniped.dev/token-exchange:cluster-001. This will prevent JWTAuthenticator configs from rejecting kubectl based logins. To use a new dynamic OAuth client, the JWTAuthenticator config will have to be updated to set .spec.audience to client.oauth.pinniped.dev/token-exchange:cluster-001 instead of cluster-001 (or an additional JWTAuthenticator can be created with the prefixed audience). This will be a transient workaround - at some future release the cluster-001 audience will no longer be automatically included for the pinniped-cli client. The token credential request API will be updated to issue a warning when the cluster-001 audience is used when the ID token is issued by a supervisor (this can be detected via the OIDC discovery document). To limit the chance that cluster-001 results in audience confusion during the transition period, the token exchange API would not include it for the pinniped-cli when a dynamic OAuth client is defined with that name (or even more strictly, if any dynamic OAuth client is defined - i.e. using the new dynamic OAuth client feature would force JWTAuthenticator config updates - we can detect if any dynamic OAuth client is defined via a live list call with limit=1).

Option 2: Require a static, reserved prefix for all dynamic OAuth client IDs

When a dynamic client is configured, the full name will not be configurable. Instead, only the suffix will be configurable and the suffix will not be allowed to contain --. The prefix will be always be client.oauth.pinniped.dev--. Thus if the suffix is set to fancy-client, the ID for the client will be client.oauth.pinniped.dev--fancy-client. The token exchange API will be updated to reject attempts to request an audience with this prefix. It may be prudent to rename the CLI OAuth client to client.oauth.pinniped.dev--pinniped-cli and prevent a token exchange for the pinniped-cli audience. JWTAuthenticator will be updated to reject audiences with the client.oauth.pinniped.dev-- prefix as well as pinniped-cli to disallow use of ID tokens issued during the authorization flow. Users will need to understand this prefixing rule as they will need to configure their app to use client.oauth.pinniped.dev--fancy-client as the client ID. Renaming the pinniped-cli would be phased effort with the old name being supported for some number of releases. The supervisor would issue warnings to indicate that kubeconfigs need to be regenerated to use the new client name (dynamic clients would not be allowed to use this reserved name).

Option 2.5: Some mix of Option 1 and 2

It is likely possible to use both options in conjunction. Note that mixing old and new versions of the supervisor and concierge may still result in audience confusion as components may not enforce all checks.

Option 3: Break apart signing keys for dynamic OAuth clients

Cryptographic isolation of signing keys is a guaranteed way to prevent any form of confusion between ID tokens issued by the authorization code flow and and ID tokens issued by the token exchange API. At a high level, we could separate the single OIDC issuer into as many as four distinct issuers (authorization code vs. token exchange x static vs. dynamic OAuth client). Consumers would be configured to trust the exact sub-issuer that they are expected to understand. The OIDC discovery protocol makes this incredibly painful to express. This technical specifics for this option have been mostly omitted as it does not seem viable.

Option 4: Migrate to token webhook

The foundational reason that audience confusion is even possible is because pinniped misuses the audience field to represent a cluster ID and an OAuth client ID at the same time. The ID token returned from the authorization code flow has no semantic relationship with the ID token returned by the token exchange API, but because they use the exact same structure and signing key, it is possible to get them confused. Options 1 and 2 above are simply attempts to slice up the audience field in a way that the two use cases are less likely to collide.

To remove the problem altogether, pinniped should move to a webhook based model for the token exchange API that uses semantically opaque tokens. These tokens will have no meaning outside of the context of the webhook and token exchange API. In particular, they will have meaning in an OAuth/OIDC context.

We will define a new requested_token_type for this purpose: urn:pinniped:params:oauth:token-type:cluster_token. The existing support for urn:ietf:params:oauth:token-type:jwt will be retained (at least for a while) but will explicitly be restricted to the pinniped-cli OAuth client. Existing JWTAuthenticator semantics for the pinniped-cli will be retained for some period of time as well. At some future point, support for urn:ietf:params:oauth:token-type:jwt will be dropped.

When the cluster_token type is used (by any client, including pinniped-cli), an opaque token will be returned after successful validation. To avoid excessive writes to etcd, this opaque token will be a short lived encrypted JWT (JWE) that will encode the information necessary to validate it as a cluster scoped token. The key used to encrypt this token will be a distinct key that is used only for this single purpose and it will not be published for external consumption. Some claims that would be relevant to include in the JWE: expiration, cluster ID, client ID, token version, username, groups, etc. The webhook would implement the v1 TokenReview API. This endpoint would be unauthenticated as it will return the same information we encode today into the token exchange ID token (after validating that the token is encrypted using the correct key, it is not expired, the client still exists, the cluster ID matches, etc). If it makes migrations easier, the webhook could also support validating ID tokens issued for the pinniped-cli.

The WebhookAuthenticator config will simply set the .spec.endpoint to https://<supervisor_url>/cluster_token_v1alpha1/<cluster_id> (along with the CA bundle if needed). Due to limitations of the upstream webhook code that we use today, we cannot encode query parameters into the URL. The v1.24 upstream changes to migrate the token webhook code to use a rest.Config as input may let us work around this limitation (which will make it easier to add extra parameters in the future). Configuring the WebhookAuthenticator will be mandatory to use the dynamic OAuth client feature, and initially it will be optional for the pinniped-cli. Over time, we will migrate to fully using only the webhook based approach.

This approach is the preferred method for solving the audience confusion problem.

Preventing web apps from caching identities without re-validation

Even with opaque tokens, once a web app learns the identity of a user, it is free to ignore the expiration on a short lived token and cache that identity indefinitely. Thus at a minimum, we should provide guidance to web apps to continue to perform the refresh grant at regular intervals to allow for group memberships to be updated, sessions to be revoked, etc.

Usability Considerations

Some considerations were mentioned already in the API section above.

Documentation Considerations

The new CRD's API will be documented, along with any other changes to related CRDs.

A new documentation page could be provided to show an example of using these new features to setup auth for a webapp, if desired.

Other Approaches Considered

• Instead of using a new CRD, the clients could be configured in the Supervisor's static ConfigMap.
  • Advantages:
    • Slightly less development effort because we wouldn't need a controller to watch the new CRD.
  • Disadvantages:
    • It would require restarting the pods upon each change, which is extra work and could be disruptive to end users if not done well.
    • Harder to integration test because it would be harder for the tests to dynamically configure and reconfigure clients.
    • Validation failures during Supervisor startup could prevent the Supervisor from starting, which would make the cost of typos very high.
    • Harder to use for the admin user compared to CRDs.

Open Questions

• Which of the options presented above should we choose? Are there other options to consider?
  • The author of this proposal doc would like to recommend that we choose:
    • "Option 1: Providing client secrets already hashed", because of the trade-offs of advantages and disadvantages discussed above. And also because client secrets should be decided by admins (see paragraph about that above).
    • And "Option 1: Explicitly associate specific clients with issuers", because it sets us up nicely for the upcoming multiple IDP feature. However, if the team does not have an appetite for doing this now, then we could choose "Option 2: Implicitly associate all clients with all issuers" for now and then reconsider when we implement the upcoming multiple IDPs feature. The security concerns raised above with Option 2 are especially important with multiple IDP support.
• Should we make the initial ID token from an authorization flow more inert? (See the audience confusion section above for more details.)

Answered Questions

None yet.

Implementation Plan

The maintainers will implement these features. It might fit into one PR.

Implementation PRs

This section is a placeholder to list the PRs that implement this proposal. This section should be left empty until after the proposal is approved. After implementation, the proposal can be updated to list related implementation PRs.