Internet-Draft Workload Identity April 2024
Hofmann & Tschofenig Expires 7 October 2024 [Page]
Intended Status:
Standards Track
B. Hofmann
H. Tschofenig

Best Current Practice for Workload Identity


The use of the OAuth 2.0 framework for container orchestration systems poses a challenge as managing secrets, such as client_id and client_secret, can be complex and error-prone. "Service account token volume projection", a term introduced by Kubernetes, provides a way of injecting JSON Web Tokens (JWTs) to workloads.

This document specifies the use of JWTs for client credentials in container orchestration systems to improve interoperability in orchestration systems, to reduce complexity for developers, and motivates authorization server to support RFC 7523.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 7 October 2024.

Table of Contents

1. Introduction

In workload scenarios dedicated management entities, also called "control plane" entities, are used to start, monitor and stop workloads dynamically. These workloads typically run micro services that interact with each other and with other entities on the corporate network or on the Internet. When one workload, acting as an OAuth client, wants to gain access to a protected resource hosted on another workload or on the Internet (referred here generically as a resource server) then authorization is typically required.

OAuth has been designed to offer help in scenarios where access to protected resources needs to be managed dynamically in a distributed system.

Each workload instance has to be provisioned with unique credentials. However, these credentials have to be configured prior and are then attached to the workload. In addition, these credentials do not have an automated rotation mechanism and are valid for an unspecified amount of time.

This requires manual configuration effort and the missing automated rotation mechanism introduce inconvenience and increase the attack surface.

"Service account token volume projection" is a feature of container orchestration systems that allows users to create JSON Web Tokens (JWTs) for their workloads. These JWTs, referred as Service Account Tokens, can be used as client credentials, as specified in RFC 7523 [RFC7523]. As these tokens are managed by the "control plane" and simply mounted to the filesystem, a workload just needs to consume this file and present it to the authorization server. In addition, service account token volume projection allows an expiration time on these JWTs to be set, allowing automated rotation of these credentials. Finally, the private key for signing these tokens is managed by the "control plane", hence removing the manual effort of configuring the client_id and client_secret.

However, there is currently no standardized way to manage these Service Account Tokens across container orchestrators. This leads to inconsistencies, and additional effort for developers as they need to support different client authentication mechanisms. In the worst case, this approach is ignored in favor of client_id and client_secret.

This specification specifies the use of Service Account Tokens in container orchestration systems, which provides a secure and scalable way to create and manage these tokens, and ensures interoperability with existing OAuth-based authorization systems.

When OAuth is used as part of the control plane entities, a Service Account Token is provisioned to the workload via the Agent. This interaction is shown in Figure 1.

To distinguish the entities, we use the term "Control Plane" to refer to the OAuth 2.0 Authorization Server that is part of the cluster's control plane. Since there are also two access tokens in play, we use the term "Service Account Token" to refer to the token issued by the Control Plane and thereby distinguish it from the other access token issued to an OAuth 2.0 client running inside the workload by the second authorization server.

It is important to note that the workload does not use the Service Account Token with resource servers directly but instead obtains access tokens from this second authorization server. To obtain these access tokens, the OAuth 2.0 client running in the workload uses the JWT client authentication grant, as defined in [RFC7523], with the Service Account Token as input. The obtained access token may be a bearer token, or a proof-of-possession token.

Figure 1 illustrates the interaction in the architecture graphically.

                      |               |
                      | Authorization |
                      | Server        |
                      |               |
                             ^ |
                             | |
  |Cluster                   | | OAuth        |
  |                          | | Exchange     |
  | +---------------+        | | to obtain    |
  | |               |        | | access token |
  | | Control Plane |        | | using        |
  | |               |        | | Service      |
  | +---------------+        | | Account      |
  |         ^|               | | Token        |
  |         ||               | v        +-----+
  |         ||            +----------+  |
  |         ||            |          |+ |        +----------+
  |  Obtain ||            | Workload || |        |          |
  |  Service||            |          ||<-------->| Resource |
  |  Account||            +----------+| | Access | Server   |
  |  Token  ||             +----------+ | Token  |          |
  |         ||                  ^       |        +----------+
  |         ||    Start Workload:       |
  |         ||     with Service :       |
  |         ||    Account Token :       |
  |         ||                  v       |
  |         ||              +-------+   |
  |         |+------------->|       |   |
  |         +---------------| Agent |   |
  |                         |       |   |
  |                         +-------+   |
  |                                     |
Figure 1: Protocol Interaction.

In Section 3 we provide more details about how the content of the tokens and the offered security properties.

2. Terminology

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

3. Architecture and Recommendations

This specification relies on the use of OAuth 2.0 [RFC6749] and [RFC7523] for client authentication using a JWT.

Service Account Tokens used in container orchestration systems are vulnerable to different types of threats, as shown in this list:

  1. Token theft: Tokens can be stolen by attackers who have already gained access to a workload. These attackers can then use these tokens to impersonate the workload and gain access to resources they should not have access to.

  2. Token reuse: Tokens can be reused by attackers to gain access to the system. However, expiration times limited the token reuse time.

  3. Misconfigured service accounts: Similar to misconfigured access to secrets, misconfigured service accounts can lead to applications gaining more privileges then necessary.

  4. Theft of token signing key: The token signing key can be stolen by attackers who have already gained access to the control plane. However, such attackers also have access to all secrets in the container orchestration system. Hence, resulting in the same impact for use of client_id and client_secret compared to using Service Account Tokens.

The following fields are populated in the Service Account Token:

  1. The 'iss' claim MUST contain a string identifying the container orchestrator.

  2. The 'sub' claim MUST contain a string identifying the workload. This is also the client_id according to [RFC7523].

  3. The 'aud' claim MUST identify one or multiple authorization servers that are intended recipients of the Service Account Token for client authorization.

Further processing requirements are specified in [RFC7523].

4. Security Considerations

This entire document is about security.

5. IANA Considerations

This document does not require actions by IANA.

6. Acknowledgements

Add your name here.

7. References

7.1. Normative References

Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <>.
Hardt, D., Ed., "The OAuth 2.0 Authorization Framework", RFC 6749, DOI 10.17487/RFC6749, , <>.
Jones, M., Campbell, B., and C. Mortimore, "JSON Web Token (JWT) Profile for OAuth 2.0 Client Authentication and Authorization Grants", RFC 7523, DOI 10.17487/RFC7523, , <>.
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <>.

7.2. Informative References

Sakimura, N., Bradley, J., Jones, M., de Medeiros, B., and C. Mortimore, "OpenID Connect Core 1.0 incorporating errata set 1", , <>.

Appendix A. Example

The functionality described in this specification can be verified using Kubernetes. Modern version of Kubernetes implement service account token volume projection, which enables the ability to inject the Service Account Token with a specific issuer and audience into the workload.

A most important parts of the configuration are (which can be found at the end of the full configuration):

  1. the path, where the application can find the token, as a file

  2. the expiration of the token in seconds

  3. the audience, which will be in the Service Account Token

  path: token
  expirationSeconds: 7200
  audience: "https://localhost:5001/connect/token"

The full configuration is shown below:

apiVersion: apps/v1
kind: Deployment
  name: simpleapp
    app: simpleapp
  replicas: 1
      app: simpleapp
        app: simpleapp
        - name: container1
          image: curlimages/curl:8.2.1
          imagePullPolicy: Always
            - sleep
            - "3600"
            - name: TOKEN_PATH
              value: '/var/run/secrets/other_token/token/token'
            - mountPath: '/var/run/secrets/other_token/token'
              name: other-token-path
        - name: other-token-path
            - serviceAccountToken:
                path: token
                expirationSeconds: 7200
                audience: "https://localhost:5001/connect/token"

The most important parts of the token, which the workload will obtain, looks as follows:

  "aud": [
  "exp": 1691752299,
  "iss": "https://kubernetes.default.svc.cluster.local",
  "sub": "system:serviceaccount:test:default"

A complete token example obtained by the workload is shown below.

  "aud": [
  "exp": 1691752299,
  "iat": 1691745099,
  "iss": "https://kubernetes.default.svc.cluster.local",
  "": {
    "namespace": "test",
    "pod": {
      "name": "simpleapp-5d7dcf96df-n7csk",
      "uid": "9fc443d7-5c7a-48d5-9679-0ee03b17d4c5"
    "serviceaccount": {
      "name": "default",
      "uid": "0bea3006-fb60-49a3-bc80-7e6884d378ae"
  "nbf": 1691745099,
  "sub": "system:serviceaccount:test:default"

To enable the authorization server to use the Service Account Token for client authentication the following configuration is needed:

  1. the client id is set to system:serviceaccount:test:default. In our case we are using the default service account in the test namespace.

  2. the public key of the token signing key. This can be either configured manually, or dynamically by referencing the JWK endpoint Kubernetes exposes, which is https://kubernetes.default.svc.cluster.local/openid/v1/jwks

Note: Authorization servers that follow the OpenID Connect Core specification, which profiles RFC 7523, will unfortunately run into problem. Here is the why.

For JWT-based client authentication [OIDC] specifies the following:

  1. The 'jti' claim is mandated for client authentication.

  2. The 'iss' claim must match the 'sub' claim. Since Kubernetes issues the tokens, and not the workload, the two do not match.

[RFC7523], on the other hand, does not mandate the use of a 'jti' claim and does not mandate that the 'iss' claim equals the 'sub' claim.

Authors' Addresses

Benedikt Hofmann
Hannes Tschofenig