This article is more than one year old. Older articles may contain outdated content. Check that the information in the page has not become incorrect since its publication.
May 2022 Security Announcement
The Flux Team has found three security vulnerabilities in Flux, and we strongly advise you to upgrade your clusters as soon as you can.
|Improper kubeconfig validation allows arbitrary code execution
< 0.29.0 >= v0.1.0
|Improper path handling in Kustomization files allows path traversal
|Improper path handling in Kustomization files allows for denial of service
< v0.29.0 >= v0.19.0
If you cannot immediately update or are hard pressed for time and need a work-around for now, please see the CVE advisories linked above for more information.
Last week the Flux Security Team disclosed three new vulnerabilities which affect v0.28 and older versions, and have a greater impact in multi-tenancy deployments.
The reason why the impact is greater in multi-tenancy deployments is due to the way that Flux/GitOps works. Flux tends to operate like a cluster admin, having permissions to apply any changes to a cluster, regardless of their scope - at namespace or cluster levels. Users having access to a source repository, or simply having access to create/alter Flux objects within a cluster can instruct Flux to apply such changes, which in single tenancy effectively means that such users have cluster admin permissions to the target clusters. The caveat being that using Flux you can add additional security controls between the user and the target cluster, for example, before each change is merged into a repository Pull Requests must be created requiring peer reviews. The Open GitOps community started defining and codifying this further - have a look at this blog post if you want to know more.
In multi-tenant environments, users with similar permissions can only affect part of a cluster, or an isolated cluster in a group of clusters. Therefore, if a user can gain escalated privileges (or deny service) at a Flux level, it will have an impact way larger than on single-tenant clusters, as the users can impact more than just themselves.
At time of writing, we are unaware of any public exploits in the wild, and therefore have no reason to believe such vulnerabilities have been actively exploited.
The advisories in detail
CVE-2022-24817 - Kubeconfig Validation
At the beginning of the Flux2’s journey we implemented a feature to apply state to remote clusters. It enables users to have a management cluster in which Flux is installed, which then applies changes to other clusters, making it ideal for some multi-tenancy scenarios in which high isolation across tenants is needed.
The connection between the management cluster and the target clusters is
done by referencing a Kubernetes secret containing a kubeconfig, which is set at the
spec.KubeConfig field in
One interesting thing about kubeconfigs is that they are quite
extensible. You can for example define an executable which will then be
automatically called by
kubectl every time it requires a new on-demand
access token. An example of this in action is
which enables AWS users to authenticate against AWS and then use the
returning JWT tokens to access their EKS clusters. Once the token expires,
the process happens again. All that is managed by
kubectl behind the scenes
without user intervention.
The problem here is that first, the use of executables in kubeconfigs was enabled by default. Meaning that a malicious tenant would be able to craft a malicious kubeconfig which could lead to privilege escalation within the cluster.
As a solution, we decided to disable this feature by default. It can
still be enabled at a cluster level via a new flag
--insecure-kubeconfig-exec being sent to the controller binary.
For cluster admins considering this feature, we also recommend the use of AppArmor and SELinux profiles to enforce at Kernel level what binaries could be executed.
CVE-2022-24877 - Kustomization Path Traversal
Flux allows users to lean on Kustomize features to make their lives easier as they go on about declaring the state of their clusters. Some of those features could result in sensitive data from the pod filesystem to be exposed into the target cluster, which could lead to a malicious tenant being privy to anything sensitive that may exist in the controller’s filesystem or attached volumes (e.g. token).
The mitigation for the path traversal was to create stronger bounds,
enforcing that all
kustomize operations happen within such bounds or
result in error.
CVE-2022-24878 - Kustomization Denial of Service
Whilst working on the mitigation of the previous CVE, we have noticed
that in some scenarios a specially crafted
kustomization.yaml could lead
kustomize-controller to enter into an endless loop, and finally
For single-tenant clusters, this would mean that an user may make a mistake and the controller stops working, potentially resulting in future reconciliations not being applied. For multi-tenant clusters, depending on the deployment model, a tenant could cause a disruption to affect not only itself but also the other tenants and potentially even the management cluster.
The solution mitigating this vulnerability was to further improve our validation and ensure that such scenarios are not processed in the first place.
Inspecting your Flux system version
To check if your system is currently vulnerable, run
flux version --context=my-cluster
--context set to the cluster you want to inspect. This will
report the current Flux binary and controller versions.
Vulnerable Flux system
To find out if your system could be vulnerable, simply find out the version of Flux. Here it’s important to check you are running all of these:
You can do this like so:
$ flux version
Updating your vulnerable system
💥 If you find the controllers versioned within the range
mentioned above, follow the upgrade procedure for your system. For
flux bootstrap, this can be done by running the command again with
the same arguments as used during install.
⚠️ Please note that if you are upgrading from below one of the versions in the following list, there are breaking changes and, pre- and/or post-upgrade notes you need to take into account: 0.29, 0.28, 0.27, 0.26, 0.24 - 0.21.
Up-to-date Flux system
An up to date Flux system should at least have versions listed below:
So in practice the output could look like this:
$ flux version
We encourage all users to keep Flux up-to-date. We offer a GitHub Action with which you can automate the Flux upgrades in a GitOps manner, without having to connect from CI to the cluster’s API, as Flux is capable of upgrading itself from Git.
Flux Security more generally speaking
It is no secret that the Flux community has been investing in security for a long time. Early on the Flux journey we re-architectured the codebase to avoid shelling out to binaries to decrease the likelihood of code execution vulnerabilities. The story about our Git integration we wrote down here.
As Security is such a central pillar of Flux, we are keen to write about it and tell you how you can benefit from all the individual features and improvements we worked on, e.g. SBOMs, CI Checks, Branch Protection, restricted pod security standard and more. Since the beginning we worked hard to ensure that we ship code that does what it needs to do, even when that means having to rewrite parts of upstream dependencies.
When we had our first security audit last year, the results were quite reassuring as most of the findings were quite small, with the exception of a RCE in kustomize-controller, which speaks to the security improvements we have been investing on, and how they are resulting in better development practices.
Another recommendation from the auditors was to implement and follow a stricter and more elaborate RFC process, which is what we did. They also recommended we get in touch with other security teams or auditors for getting feedback on a refined and more general multi-tenancy proposal - which we also did as mentioned below.
What’s next for Flux
The way we shaped the Flux Roadmap for GA was
- Feature party with Flux Legacy
- Stable APIs (this was after the controller refactoring which consolidated functionality in fluxcd/pkg)
- Straightforward multi-tenancy implementation
- GA release
Here is the status quo regarding multi-tenancy:
Flux2 supports multi-tenancy, and users have been using it in production for some time now.
The documentation around the subject covers a bootstrap example to help users kick start their multi-tenancy deployments. And also how to implement control plane isolation with the multi-tenancy-lockdown.
In summary, the documentation needs expanding to better inform users around the security risks of multi-tenancy and the recommended deployment models for their specific isolation/security requirements.
There are proposed changes that would further improve Flux in multi-tenancy environments, by for example enabling tenants to share resources amongst themselves. Such changes must be progressed once the security impact of such changes have been assessed.
To help us get this right, we are engaging with the CNCF TAG Security. This is the upstream group where key contributors and experts of the CNCF Landscape assemble and define security best practices across all the individual Cloud Native projects. We are asking them for an independent security review and recommendations, particularly around multi-tenancy.
If you want to join the conversation, we are all ears. Please refer to the open RFC documents and have your say there. We definitely want to get this right for everyone.
- RFC: Define Flux tenancy models #2086
- RFC: Access control for cross-namespace source refs #2092
- RFC: Flux Multi-Tenancy Mode #2093
In addition to that we are working hard to round up features, improve their performance, security and overall stability.
If you want to follow all the other GA related work, we explained how to do that here and if you would like to participate in any of the discussions, come and find us on Slack or in our regular meetings. We are always looking forward to growing Team Flux and the closer we get to GA, it’s getting even more important to have all voices heard.