Seven Kyverno CVEs were published by NVD on November 5, 2026, all assigned by VulnCheck and all fixed by the same release: Kyverno 1.19.2. Six are rated 9.8 or 9.9 in CVSS 3.1, the seventh 9.1.
The common thread runs through almost all of them: Kyverno lets users confined to a namespace write policies, then executes those policies with its own service accounts, which hold cluster-wide rights. Six of the seven flaws turn that gap into cluster takeover or writes into other tenants' namespaces. The seventh bypasses image signature verification.
If you run a cluster shared between teams with Kyverno as the policy engine, this is precisely the model these flaws break.
The Seven Vulnerabilities
| CVE | Affected versions | CVSS 3.1 (VulnCheck) | CVSS 4.0 (VulnCheck) | CWE | Mechanism |
|---|---|---|---|---|---|
| CVE-2026-122302 | before 1.19.2 | 9.8 | 9.3 | CWE-269 | generate rules: checks the controller's rights, not the author's |
| CVE-2026-122303 | 1.17.0 up to 1.19.2 (excluded) | 9.9 | 9.4 | CWE-862 | NamespacedMutatingPolicy: fabricated target in another namespace |
| CVE-2026-122304 | 1.17.0 up to 1.19.2 (excluded) | 9.9 | 9.4 | CWE-863 | CleanupPolicy: context loaded with no policy, Secrets readable |
| CVE-2026-122310 | 1.19.1 only | 9.9 | 9.4 | CWE-436 | apiCall.urlPath validated cleaned, executed raw |
| CVE-2026-122311 | 1.16.0 up to 1.19.2 (excluded) | 9.9 | 9.4 | CWE-22 | resource.get() / resource.post(): .. segments |
| CVE-2026-122313 | 1.17.0 up to 1.19.2 (excluded) | 9.9 | 9.4 | CWE-269 | NamespacedMutatingPolicy: edits another tenant's RoleBinding |
| CVE-2026-122321 | 1.18.0 up to 1.19.2 (excluded) | 9.1 | 8.5 (HIGH) | CWE-367 | verifyImages: cache keyed by tag, not digest |
All were published on 2026-11-05. NVD status: Received: NVD hasn't analyzed them yet, the scores are VulnCheck's, and no CPE configuration is published. The CVSS 3.1 vectors are AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H for the five rated 9.9, AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H for CVE-2026-122302 and AV:N/AC:L/PR:L/UI:N/S:C/C:L/I:H/A:L for CVE-2026-122321.
The Root Problem: Who Runs the Policy?
Kyverno runs several controllers (admission, background, cleanup), each with a broadly privileged ServiceAccount. When a namespace admin creates a policy, the controller is what applies it. If Kyverno doesn't check that the policy stays within its author's scope, the author borrows the controller's rights.
Generate Rules (CVE-2026-122302)
According to NVD, the admission check for generate rules verifies the controller's permissions, not those of the policy author. A namespace admin can therefore write a namespaced Policy whose generate rules target other namespaces or cluster-scoped kinds, and the background controller creates or clones them on the admin's behalf.
Namespaced Mutating Policies (CVE-2026-122303 and CVE-2026-122313)
Two flaws hit NamespacedMutatingPolicy and its CEL expression targetMatchConstraints.expression, which selects the resources to mutate. The expression's output is not checked against the policy's namespace:
- CVE-2026-122313: the expression emits a literal target, such as another tenant's RoleBinding, which the background controller mutates to add the attacker's subjects. Roles, NetworkPolicies and ConfigMaps are also named;
- CVE-2026-122303: in mutate-existing mode, the expression returns a fabricated CEL map with an attacker-chosen
metadata.namespace, and the controller overwrites Secrets, Deployments or policies cluster-wide.
Cleanup Policies (CVE-2026-122304)
The cleanup controller builds its context loader with no associated policy (a nil policy, per NVD). A namespaced CleanupPolicy with apiCall or globalReference context entries can then read Secrets from other namespaces and modify validatingwebhookconfigurations, using the controller's ServiceAccount. Tampering with validating webhooks can mean switching off admission controls altogether.
Two Path Traversals (CVE-2026-122310 and CVE-2026-122311)
- CVE-2026-122311: the CEL functions
resource.get()andresource.post()don't validate path segments. A tenant who can create aNamespacedValidatingPolicyinserts..and reads or writes any cluster object with the admission controller's ServiceAccount. NVD calls it cluster takeover; - CVE-2026-122310, specific to 1.19.1: Kyverno validates a
Policy'scontext.apiCall.urlPathafter cleaning it, but executes the raw path. Since the Kubernetes client drops whatever follows#or?,../segments hidden behind those characters reach cluster-scoped endpoints. Outcome per NVD: cluster admin.
A textbook CWE-436 case: the validator and the executor don't interpret the same string the same way.
The Image Verification Cache (CVE-2026-122321)
This one isn't about namespace isolation but about the supply chain. verifyImages caches signature verification results keyed by the image reference (a tag, for example) rather than the digest actually verified. An attacker who can submit workloads and move a tag in the registry:
- gets a signed image verified under that tag, warming the cache;
- repoints the tag to an unsigned image;
- submits a workload: the cache says "verified", signature and attestation checks are skipped, and
mutateDigestpins the unverified digest in the manifest.
The admitted image then carries a pinned digest that looks like a controlled deployment.
Affected and Fixed Versions
| Branch | Status |
|---|---|
| Before 1.16.0 | affected by CVE-2026-122302 |
| 1.16.x | CVE-2026-122302, CVE-2026-122311 |
| 1.17.x | + CVE-2026-122303, CVE-2026-122304, CVE-2026-122313 |
| 1.18.x and 1.19.0 | + CVE-2026-122321 |
| 1.19.1 | all seven, including CVE-2026-122310 |
| 1.19.2 | first version outside the range for all seven |
⚠️ These bounds come from the NVD descriptions ("before 1.19.2"), not from a CPE configuration. Check the recommended version, and any maintenance releases on other branches, in the project's GitHub advisories (GHSA-q62c-r9v2-cf64, GHSA-jcw5-3wcg-fqx5, GHSA-p58h-r4j3-h5vx, GHSA-rcxg-3f92-3mgr, GHSA-3xfv-vq8x-qhg4, GHSA-4pqm-wqq6-h9q2, GHSA-f96w-622v-rqpq, GHSA-2r43-6962-xcw5).
Check the deployed version:
kubectl get deploy -n kyverno -o jsonpath='{range .items[*]}{.metadata.name}{" "}{.spec.template.spec.containers[0].image}{"\n"}{end}'
Who Is Actually Exposed
Going by their descriptions, every one of these flaws requires an authenticated user on the cluster who can create certain Kyverno resources or submit workloads (VulnCheck's vector for CVE-2026-122302 nonetheless says PR:N, while its description talks about namespace admins). Your risk depends on your model:
- Multi-tenant cluster where teams or customers create their own
Policy,CleanupPolicy,NamespacedValidatingPolicyorNamespacedMutatingPolicy: maximum exposure, it's the scenario every CVE describes; - Cluster where only platform admins write policies: much smaller surface for the six isolation flaws, but CVE-2026-122321 still matters as soon as someone can submit workloads and change a tag in the registry;
- Compromised service account in a namespace: it becomes a springboard to the whole cluster if its role allows creating Kyverno policies.
Detection
- Inventory namespaced policies and who wrote them:
kubectl api-resources | grep -i kyverno
kubectl get policies.kyverno.io -A
kubectl get cleanuppolicies -A
generaterules targeting another namespace or a cluster-scoped resourcetargetMatchConstraintsexpressions returning a literal namespace other than the policy's ownapiCall.urlPathcontaining..,#or?, andresource.get()/resource.post()calls with..segments- RoleBindings, Secrets and
validatingwebhookconfigurationsmodified by a Kyverno ServiceAccount for no obvious reason: the Kubernetes API audit logs are where this shows up - Admitted images whose digest matches no known signature
Mitigation
1. Upgrade to Kyverno 1.19.2
It's the only single answer to all seven.
2. Meanwhile, restrict who can write policies
Remove create / update rights on Kyverno policy resources (Policy, CleanupPolicy, NamespacedValidatingPolicy, NamespacedMutatingPolicy) from namespace users. It hurts a self-service model, but it closes the door shared by six of the seven flaws.
3. For image verification
Reference images by digest rather than by tag in your manifests, and turn on tag immutability in the registry where available. A tag that can't move defeats the CVE-2026-122321 scenario.
4. After the upgrade
Malicious policies and the changes they already made don't go away with the patch. Review RoleBindings, Secrets and validating webhooks, and rotate any secrets that may have been read.
For a broader take on container security, see Docker and Kubernetes: Critical CVEs and Container Security.
Why Continuously Monitor Your Kubernetes Platform
A policy engine is supposed to be the layer that keeps teams from spilling into each other. Seven critical CVEs published on the same day, with no NVD analysis and no CPE, is exactly the case where a product search in the usual databases returns nothing for days.
With cveo.tech, declare Kyverno and the other building blocks of your platform with their exact versions, and get an alert as soon as a CVE affects them, even before NVD has analyzed it.