Tutorials

Ingress-NGINX CVE-2026-116472: Why a Low CVSS Can Cause a Shared Outage

A Kubernetes advisory exposes a tenant-to-controller trust boundary: inspect mirror-target permissions, limit annotations, and plan migration from retired ingress-nginx.

Approximately 5 min read

A Kubernetes security advisory published November 5, 2026 describes a counterintuitive failure mode: an actor who can create an ordinary Ingress resource may be able to disrupt the data plane of a shared ingress-nginx controller. The issue, CVE-2026-116472, carries a CVSS 3.1 score of 2.7 (Low). Yet its potential blast radius is not limited to the actor’s own application. This is a source-led configuration and trust-boundary analysis, not an independent exploit reproduction.

The upstream Kubernetes issue states that a missing check permits the mirror-target annotation to reach an unauthenticated configuration API inside the ingress-nginx controller. That API can receive writes that corrupt the controller’s live configuration. The result can be dropped traffic for other Ingresses served by the same controller until a subsequent Ingress change or controller restart restores the data. The advisory also describes a theoretical traffic-redirection consequence, while explicitly downplaying its incremental practical risk.

The important boundary is not namespace isolation

Kubernetes RBAC can authorize a team to create Ingress resources in its namespace without granting it cluster-administrator rights. That authorization does not necessarily mean every annotation accepted by the ingress controller has namespace-local effects. An Ingress is a declarative API object; the controller interprets its metadata and translates it into behavior of a shared proxy.

The distinction is critical. The API server checks whether the actor may write the Ingress object. The controller must separately decide whether the requested behavior is safe to execute. A high-impact annotation can cross the boundary between a tenant-owned object and shared routing infrastructure. In this CVE, the dangerous bridge is the mirroring target’s ability to address an internal configuration endpoint.

This explains why the low CVSS score and the operational concern can coexist. The published vector requires high privileges in the CVSS model, which sharply limits the base score. A platform with broad delegated Ingress-write permissions, however, may have a larger set of actors who meet that prerequisite than its administrators expect. CVSS describes a generic vulnerability; it does not model the number of tenant teams or the availability impact of a particular organization’s shared ingress topology.

Establish whether the controller exists

The Kubernetes Security Response Committee advisory says clusters without ingress-nginx are not affected by this issue. A first inventory check is:

kubectl get pods --all-namespaces --selector app.kubernetes.io/name=ingress-nginx

A negative result is a useful starting point, not a complete inventory guarantee: custom labels or alternative installation layouts can make selector-based discovery incomplete. Cross-check installed Helm releases, controller deployments, image names and the IngressClass objects that route traffic. Distinguish the community ingress-nginx controller from unrelated NGINX-based ingress products; a similar product name does not establish identical exposure.

For every ingress-nginx instance, identify which namespaces and identities can create or modify Ingresses it processes. Then determine whether that instance handles only one trust domain or aggregates unrelated applications. The latter makes a configuration-level outage more consequential.

Two immediate mitigations, and what each does

The upstream advisory names two mitigations. One is an admission policy that rejects Ingress objects specifying the mirror-target annotation. The other is setting the controller ConfigMap parameter annotations-risk-level to Medium or Low. The latter restricts a broader class of risky annotations; it may therefore break legitimate mirroring or other advanced routing configurations. Test the change against the actual inventory rather than assuming it is behavior-preserving.

An admission rule is narrower but must cover both creation and updates. It should be enforced at the API admission boundary for all namespaces handled by the controller, not merely documented as a team convention. Where policy engines differ, test the exact annotation key and rejection behavior in a staging cluster. A policy that blocks new objects but ignores updates leaves a straightforward gap.

Neither mitigation should be presented as a universal replacement for upgrading or migration. The security committee states that upstream ingress-nginx maintenance ended in March 2026, and that Kubernetes will not issue an upstream fix for this CVE. It identifies SUSE’s ingress-nginx fork version 1.15.2-prime1 as carrying a fix. That version is a vendor-fork release, not a patch version to assume exists in the archived upstream project. Verify vendor support and the precise distribution in use before planning an upgrade.

Review existing objects without exercising the flaw

Defensive detection should inspect existing Ingress manifests for mirror-target annotations, with special attention to localhost targets, which the advisory calls out as suspicious. Treat a match as an investigation lead rather than proof of compromise: legitimate traffic mirroring may exist, and the advisory does not claim every use is malicious.

For each finding, record the namespace, object owner, creation/update history, controller instance and change ticket where available. Review controller events and configuration-reload behavior around unexplained routing interruptions. An outage alone cannot establish that this CVE was exploited. Do not probe a production controller by sending crafted mirroring requests to internal endpoints: the potential impact is the very availability failure under investigation.

A practical safe validation plan is to create an ordinary Ingress with a harmless mirroring annotation in an isolated test environment and confirm that the chosen admission policy rejects it. This validates the control path without attempting to corrupt an internal API.

Migration is a security control, not just maintenance

The unusual feature of this advisory is the absence of an upstream repair path. A temporary annotation restriction may reduce this specific risk, but it does not reverse the lifecycle decision. Operators should inventory ingress-nginx dependencies, choose a maintained controller or supported fork, and rehearse migration with representative traffic.

That rehearsal should cover TLS termination, authentication redirects, rewrites, canary behavior, request-size limits, custom snippets and observability. Those features are often encoded in implementation-specific annotations and do not automatically translate to another controller. Run both configurations against comparable staging traffic, verify application-level behavior, and plan a rollback boundary before changing the public entry point.

The general lesson is larger than CVE-2026-116472: permission to submit a configuration object is not equivalent to permission to exercise every capability the controller exposes. In multi-tenant systems, annotations and extension points are privileged inputs even when the objects carrying them are namespace-scoped.

Evidence and limits

The upstream issue, security mailing-list notice and Kubernetes announcement establish the vulnerability, stated consequences, official mitigations and vendor-fork fix. This article adds a threat-boundary interpretation and defensive validation plan. It does not claim to have reproduced exploitation, measured outage duration, verified an active campaign or tested a particular vendor’s patch. Those questions require separate evidence.

Sources and further reading

Continue reading

More in Tutorials ↗