Sources checked: . FailSafe Research.
F5 released fixes on 18 November 2026 for five NGINX vulnerabilities that OX Security has grouped under the name nginx2root. Each one is described as an independent path to remote code execution when the relevant NGINX feature is enabled. They do not need to be chained. The practical question for most teams is narrower than the headline: which of these features do we actually run, on which build, and how do we prove the binary was replaced?
OX Security published the analysis and credited its Xint research team with finding all five. The same post says Wiz Research, Ikotas Labs, Metnir and Rapid7 independently reported related issues to F5. Version ranges and fix releases below come from that report and the F5 release. Confirm them against the current F5 advisory before acting.
The five CVEs and affected versions
| CVE | Component | Affected versions | Fixed in |
|---|---|---|---|
| CVE-2026-123007 | HTTP/2 flow control | 1.9.5 to 1.31.8 | 1.30.7 and 1.31.9 |
| CVE-2026-123634 | DNS resolver with mirror | 1.13.4 to 1.31.8 | 1.31.9 |
| CVE-2026-123633 | HTTP/2 with upstream proxying | 1.19.7 to 1.31.8 | 1.30.7 and 1.31.9 |
| CVE-2026-121810 | HTTP/3 with upstream proxying | 1.25.0 to 1.31.8 | 1.31.9 |
| CVE-2026-113425 | Mail proxy authentication | 0.5.15 to 1.31.8 | 1.31.9 |
NGINX Plus is affected on the 37.x line. CVE-2026-123007 is fixed in 37.0.8.1 and 37.3.0.1, with affected ranges 37.0.0 to 37.0.7.1 and 37.1.0 to 37.2.0.1. CVE-2026-123633 is fixed in 37.0.7.1 and 37.2.0.1, with affected ranges 37.0.0 to 37.0.6.1 and 37.1.0 to 37.2.0. For releases R34 to R36, the report says no fix exists and the path is to upgrade to the 37.x line.
What the report does and does not establish
Several details matter more than the word critical.
- Authentication is not required. The report describes unauthenticated attack paths, which is why the flaws are being treated seriously on internet-facing systems.
- Exposure depends on configuration and version. A server that does not enable the relevant feature is not exposed through that flaw. The set is broad because HTTP/2 and upstream proxying are common, not because every NGINX instance runs every affected path.
- One public exploit exists. The report states that Rapid7 published a working exploit for the HTTP/2 flow control flaw, CVE-2026-123007, and that no exploitation in the wild had been confirmed at publication.
- No CVSS score appears in the report. The critical label is OX Security's framing. Treat it as a prompt to check your own exposure, not as a substitute for the F5 severity and vector data.
- The name does not prove root on every host. Remote code execution is the claim. Whether a successful attack reaches root depends on the process permissions, sandboxing and surrounding controls on that system.
The HTTP/2 flaw is the one to prioritize. The report says it has been present since NGINX introduced HTTP/2 in 2015, which means a long-lived build can carry it without any recent change to explain the risk.
Exposure is a configuration question
Answer these for every NGINX deployment you operate, including the ones behind a CDN or load balancer:
| Feature | Related CVE | Question |
|---|---|---|
| HTTP/2 | CVE-2026-123007, CVE-2026-123633 | Is HTTP/2 enabled on an affected build, including through TLS configuration or a default server block? |
| Upstream proxying with HTTP/2 | CVE-2026-123633 | Does the server proxy to upstreams over HTTP/2, and which build serves that path? |
| HTTP/3 and upstream proxying | CVE-2026-121810 | Is HTTP/3 enabled, and is an upstream proxy configured on the same build? |
| DNS resolver with mirror | CVE-2026-123634 | Does the configuration use a resolver together with mirroring, or resolve upstream names during requests? |
| Mail proxy authentication | CVE-2026-113425 | Is the mail proxy module enabled and exposed, including on a host that is not primarily a web server? |
A version inventory alone is not enough. Two servers can run the same build with different exposure because one enables HTTP/3 and the other does not. Record both the build and the enabled features for each internet-facing instance.
Patch, then verify the binary changed
The most common failure after a disclosure like this is a configuration reload that leaves the vulnerable binary running. Apply the fixed package, restart the service, and confirm the running process reports the new build. For NGINX Plus, confirm the package version rather than a marketing release label.
Three environments need separate checks:
- Distribution packages. A Linux vendor may backport the fix without changing the upstream version that the nginx -v command prints. Check the full package version and the vendor advisory, not only the NGINX version string.
- Container images and base images. The fixed package has to be present in the image you deploy. A rebuilt application layer can still inherit an old NGINX base. Confirm the image digest and the package inside it.
- Managed and edge services. If a provider terminates TLS or runs the proxy for you, ask which build serves your traffic and when the fix is applied. Your own patching does not cover a provider-managed edge.
Keep the evidence: the package version before and after, the restart or rollout timestamp, the instance or image digest, and the person who confirmed it. A ticket marked resolved is not proof that the running process changed.
Interim mitigations if patching has to wait
The report describes partial mitigations for teams that cannot patch immediately. They reduce exposure; they do not replace the fixed build.
- Disabling HTTP/2 addresses only CVE-2026-123007 and CVE-2026-123633. It does not cover the other three.
- Disabling HTTP/3 addresses the HTTP/3 path, and avoiding dynamic upstream DNS resolution during requests reduces exposure to the resolver flaw.
- Removing public exposure from the mail proxy addresses the mail path. A mail proxy does not belong on an internet-facing web tier unless there is a specific reason.
If you take a temporary mitigation, write down which flaw it covers and when it will be removed. A setting added during an incident can outlive the reason for it and quietly reduce capability later.
What to test after the patch
An authorized test can confirm that the deployed service behaves as the change intended, within a scope you control. Useful checks include:
- Whether the patched build is the one answering requests, checked from the network rather than from the deployment record.
- Whether the expected protocol versions are enabled or disabled on each exposed endpoint.
- Whether configuration generated by automation matches the reviewed configuration, including server blocks added since the last review.
- Whether the mail proxy, resolver and upstream paths are reachable from where they should and from where they should not be.
- Whether monitoring records the relevant service restart and version change, so the next disclosure has a detection path.
Keep the result tied to the exact build and configuration tested. A clean result on a staging build tells you nothing about a production instance that was rebuilt from a different base image.
Where FailSafe fits
FailSafe tests authorized applications, APIs, cloud infrastructure and the systems around them. A scoped engagement can verify externally observable protocol behavior, review the configuration that governs exposure, and produce evidence that ties a result to the deployed build. The application and infrastructure testing scope describes how the boundary is agreed. The rules-of-engagement guide covers the authorization and stop conditions that should be settled before any test touches a production edge.
We do not publish exploits, payloads or reproduction steps for an unpatched flaw, and we do not test third-party infrastructure without that provider's authorization. If you suspect active exploitation, contact your incident-response provider and preserve logs before restarting or rebuilding the affected system.
What this article does not establish
This article does not confirm exploitation in the wild, a CVSS score, or the full set of affected configurations. The version ranges, fix releases and mitigation guidance are attributed to the OX Security report and the F5 release, and they can change as vendors publish corrections. Confirm the current advisory, your package versions and your provider's status before relying on any figure here.
Questions & answers
Questions about this topic
Direct answers to common questions about this page.
The OX Security report states that no exploitation in the wild had been confirmed at publication, and that a working exploit for CVE-2026-123007 had been published by Rapid7. Treat that as a fast-moving situation and check the current F5 advisory.
No. Disabling HTTP/2 addresses CVE-2026-123007 and CVE-2026-123633 only. The HTTP/3, resolver and mail proxy flaws need their own treatment, and patching is the durable fix.
No. A reload can keep the old binary running. Install the fixed package, restart the service, and confirm the running build from the network and the process.
The report lists affected and fixed versions on the 37.x line. Releases R34 to R36 have no fix in that report, and the guidance is to upgrade to 37.x. Confirm current details with F5.
Check the package version inside the image and record the image digest. Rebuilding an application layer does not update an inherited NGINX base.
Related Articles
PHP2Shell: affected PHP versions, patches, and validation
PHP2Shell response guide for CVE-2026-109618 and CVE-2026-109619: fixed PHP versions, exposure review, and evidence after patching....
BeyondTrust BT26-05: patch the appliance and investigate prior access
BeyondTrust’s October disclosure of CVE-2026-102764 describes earlier zero-day exploitation. A remote-support security and validation guide....
GitLab CVE-2026-110264: critical RCE patch and CI/CD review
GitLab's CVSS 10.0 regular-expression flaw: affected CE/EE branches, October 21 fixes, and a CI/CD security validation checklist....
Put this guidance to work
Bring the exposed endpoints, builds and configuration you want checked. FailSafe can verify the deployed behavior and produce bounded evidence.