On this page (7 sections)

Kubeflow Pipelines, the workflow engine used to build and run machine learning jobs on Kubernetes, has a critical authorization flaw in its API server. In multi-user mode, two upload endpoints skip the permission check when no namespace is supplied, which is the default. A low-privilege tenant can then publish shared pipelines or slip a new version into an existing shared one.

If you run a shared Kubeflow installation, especially one where several teams or customers have their own namespaces, treat this as a priority upgrade. The details below come from the GitHub security advisory and the CVE record.

What happened

The advisory describes an omission in the API server's HTTP upload handlers. The endpoints POST /apis/v2beta1/pipelines/upload and POST /apis/v2beta1/pipelines/upload_version rely on a helper called canUploadVersionedPipeline to decide whether the caller may write. According to the advisory, that helper returns without performing a Kubernetes SubjectAccessReview whenever the resolved namespace is an empty string.

An empty namespace is how Kubeflow Pipelines addresses shared, cluster-wide pipelines. It is also what you get when a caller simply leaves out the namespace query parameter. Because the setting REQUIRE_NAMESPACE_FOR_PIPELINES defaults to false and, per the advisory, is not set in any shipped manifest, a default multi-user install accepts these requests and reaches the unguarded branch.

The advisory points out that the gRPC code path for the same kind of write does authorize empty-namespace requests, by mapping them to the system namespace, with a comment saying shared-pipeline writes must still be authorized. The advisory concludes that the HTTP behavior is an omission, not intended behavior.

Why it matters: from upload to code execution

Uploading a pipeline is not just storing a file. A pipeline version defines the container image, command and arguments that a workload runs. The advisory says an attacker can create cluster-shared pipelines and inject versions into existing shared pipelines, replacing what other tenants execute.

The CVE record adds important limits on how far this goes:

  • An enabled, unpinned recurring run that follows the latest version can pick up the injected specification and run it in another tenant's namespace, under that workload's service account.
  • Pinned runs, or runs with an embedded specification, are not automatically substituted.
  • A low-privilege tenant can exploit the flaw directly. An unauthenticated caller can do so only if the API server is reachable without an authenticating proxy. Deployments behind such a proxy do not expose anonymous access.

In plain terms, this is a cross-tenant integrity problem. The weakness class is missing authorization: the server checks permissions in most places, but on this path it never asks whether the caller is allowed to write. The impact can reach well beyond ML metadata, because the injected content is a container specification that runs with another team's permissions.

Who is exposed

Exposure depends on several conditions:

Exposure factors for CVE-2026-77596
ConditionHigher riskLower risk
Deployment modeMulti-user modeNot multi-user mode (the advisory code returns early for non-multi-user installs, and the CVE describes the multi-user endpoints)
VersionBefore 2.18.02.18.0 or later
REQUIRE_NAMESPACE_FOR_PIPELINESUnset or false (the default)Set to true rejects empty namespaces in the quoted validation code, but the research does not call this an official fix
API server reachabilityDirectly reachable, no authenticating proxyBehind an authenticating proxy
Recurring runsEnabled and unpinned, following the latest versionPinned or embedded-specification runs
Even behind an authenticating proxy, any authenticated tenant lacking write permission on shared pipelines can still abuse a vulnerable version, because the missing check is inside the API server.

How to check your exposure

First find out which API server image you are running. This read-only command lists the container images of all pods and filters for the Kubeflow Pipelines API server image name used in the advisory. Run it with a kubeconfig that can list pods in all namespaces, and compare the tag or digest it prints against the 2.18.0 release.

List running Kubeflow Pipelines API server images (read-only, needs cluster-wide pod read access)Shell
kubectl get pods -A -o jsonpath='{range .items[*]}{.metadata.namespace}{"\t"}{.metadata.name}{"\t"}{range .spec.containers[*]}{.image}{" "}{end}{"\n"}{end}' | grep -i kfp-api-server

Next check whether anyone has set REQUIRE_NAMESPACE_FOR_PIPELINES in your deployments. This read-only command searches the YAML of all deployments for the variable. No output means it is not set in any deployment spec, so it takes the default value of false.

Search deployment specs for REQUIRE_NAMESPACE_FOR_PIPELINES (read-only)Shell
kubectl get deployments -A -o yaml | grep -n REQUIRE_NAMESPACE_FOR_PIPELINES

Then review what already exists. Look at the shared (cluster-wide, no namespace) pipelines in the Pipelines UI or API and check who created each one and when. Any shared pipeline or new version you do not recognize deserves a closer look, along with recurring runs that follow the latest version.

What to do now

  1. Upgrade Kubeflow Pipelines to 2.18.0 or later, the version the release notes and CVE record name as the fix.
  2. Confirm the running API server image matches the fixed release using the image listing command above.
  3. Make sure the API server is not directly reachable from untrusted networks. Put an authenticating proxy or equivalent access control in front of it.
  4. Audit shared pipelines and recent versions. Remove or quarantine anything unexpected, and note the creator and time of each.
  5. Review enabled recurring runs that follow the latest version. Where possible, pin them to a specific version so a new upload is not picked up automatically.
  6. Check the container images, commands and arguments in recent pipeline versions for anything you did not approve, and review the workloads they ran.
  7. If suspicious versions have already run, treat the affected namespaces and service accounts as potentially exposed and rotate credentials those workloads could reach.
Caution. Upgrading the API server changes a shared control plane component. Test the new version first and schedule it so running pipelines are not interrupted unexpectedly.
Based on the validation code quoted in the advisory, setting REQUIRE_NAMESPACE_FOR_PIPELINES to true makes the server reject empty-namespace pipeline operations. The research does not present this as an official workaround, and it could break workflows that rely on shared pipelines. Use it, if at all, only as a short stopgap while you plan the upgrade, and test it first.

Detection and monitoring afterwards

Once you have patched, keep watching the places this bug would have touched. Review Kubernetes audit logs and API server logs for upload requests to the two endpoints, particularly ones with no namespace parameter or from identities that should not publish shared pipelines. Alert on new shared pipelines and new versions of existing ones, and on recurring runs that start with an image you have not seen before.

If your Kubeflow dashboard or API is exposed through a public hostname, the web layer in front of it matters too. A website firewall or authenticating reverse proxy narrows who can reach the API at all, and certificate and DNS monitoring can show when the endpoint's public footprint changes. Neither replaces the upgrade.

Timeline and confidence

  • Oct 28, 2026: the GitHub advisory GHSA-rjr3-43j3-4q8g was published, according to its page.
  • Nov 5, 2026: the CVE record shows its publication date.
  • Nov 9, 2026: the GitHub advisory was last updated.

A caveat on sourcing: the research rests on the vendor's own GitHub advisory and the CVE record, with no independent confirmation and no reports of active exploitation. The advisory says the issue was reproduced against the real API server binary, but we have not seen third-party verification. We do not describe exploitation steps here, and you do not need them to decide to patch.

Sources

This article is based on the public sources below. Details can change as vendors and researchers publish more; check the original advisories before acting on version numbers.

  1. CVE-2026-77596 - GitHub Advisory Databasegithub.com
  • Kubeflow
  • Kubernetes
  • CVE-2026-77596
  • authorization
  • MLOps