Your NGINX version check will reassure you, right up until you find the one your cloud provider runs for you, with HTTP/2 switched on.
Imagine you’re told one of the locks on your building is faulty. Easy enough, you think, we’ll change the locks. Then you start counting the front doors. There’s the main entrance, a side door each team put in themselves, a loading bay nobody’s opened in a year, and a couple the landlord holds the keys to. And the label on each lock tells you its model, not whether someone’s already swapped the barrel.
One bug, four questions
On 18 November, F5 shipped out-of-band fixes for NGINX, including two unauthenticated remote code execution flaws in its HTTP/2 handling, CVE-2026-123007 and CVE-2026-123633. Five research teams found them independently. Wiz says CVE-2026-123007 goes back to release 1.9.5, in 2015, and showed it running code on a server that only served static files. As Wiz puts it: “Websites and APIs can be exposed even when their own application code is secure.” Xint calls the set “nginx2root”.
The fix is 1.31.9 mainline or 1.30.7 stable. That sounds simple. It isn’t, because “are we exposed?” is really four questions. Where does NGINX run? What build is it really? Is HTTP/2 switched on and reachable? And what sits underneath it?
Why the version number lies
Xint says it plainly: “The version number alone may not show whether a server is patched.” Ubuntu 22.04’s NGINX reports 1.18.0 whether it’s fixed or not, because the distro backports the fix into the same version. The real answer is in the full package version (Xint’s example compares 1.18.0-6ubuntu14 with 1.18.0-6ubuntu14.22), checked against your distro’s advisory.
Containers make it worse. The NGINX inside an image is whatever was there when the image was built, whatever the host package says. And then there are the copies that aren’t really yours. A public issue on Azure’s AKS app-routing repo has a customer asking about a managed ingress image shipping NGINX 1.27.1 with HTTP/2 on by default and, in their words, no supported way to switch it off. That’s one customer’s question, not a Microsoft statement, but it’s the right question for anyone running a managed ingress add-on.
It’s not exposed until it’s switched on and reachable
CERT-EU is clear about CVE-2026-123007: “The only precondition is HTTP/2 enabled on a listening port.” That’s a configuration fact, not a version fact. An old build with HTTP/2 off is a lower priority than a newer unfixed one serving HTTP/2 to the internet.
So rank by reachability, internet-facing first, which is what CERT-EU recommends too. Where you can’t patch straight away, set http2 off;, remove http2 from your listen directives, reload, and check the old worker processes have actually exited. Two caveats. Anything that relies on HTTP/2, such as gRPC, will notice. And CERT-EU says it only mitigates two of the five flaws, so it buys time, it doesn’t close the ticket.
The container isn’t a wall
This is the bit that turns a web server bug into an infrastructure problem. Xint chained the NGINX flaws with a Linux kernel SCTP bug: “In our tests, the chain also escaped a Docker container running nginx and obtained root on the underlying host.” Since every container on a host shares its kernel, that means the host and every container on it. They’ve demonstrated it on AWS, Azure and Google Cloud.
The kernel half had no fix at disclosure. The mitigation is a line of configuration on every host that doesn’t use SCTP, blocking the module under /etc/modprobe.d/. There’s nothing to “patch”, so someone has to apply it across the fleet, prove it’s there, and take it off when the kernel fix ships. And as we said about running versus installed kernels, what’s on disk isn’t always what’s running.
Same question, again
This landed a day after Java4Shell, a week after Apache httpd’s own HTTP/2 flaw, and alongside another PHP release. A platforms leader at a global manufacturer told us recently that their real pain is patching at machine speed, and their service providers aren’t keeping up. A CTO said something similar to us after log4j: the pain was never the patching, it was not knowing where it ran.
What to check this week
- List every NGINX. Distro packages,
nginxand OpenResty base images, ingress controllers, and products you buy that ship it. - Check the real build. Use
dpkg -l nginxorrpm -q nginxand your image build, notnginx -v, against 1.31.9, 1.30.7 or your distro’s advisory. - Check HTTP/2 from outside. Xint’s test:
openssl s_client -connect HOST:443 -servername HOST -alpn h2, and look forALPN protocol: h2. - Patch or rebuild internet-facing copies first. Where you can’t, switch HTTP/2 off and write down who turns it back on.
- Block SCTP where you don’t use it.
echo "install sctp /bin/false" | sudo tee /etc/modprobe.d/disable-sctp.conf, until the kernel fix ships. - Ask your provider, in writing. For managed ingress, get the patched image and the date it lands.
Step one decides how long everything else takes.
See it in practice
Fixing this is less about one patch and more about one change applied everywhere and then proven. For a Nordic enterprise, our engine applied and verified one security rule across 21 security groups in 13 regions in 15 minutes. That’s the shape of this job: find every copy, change it under your rules, and show it’s done.
Anyway, don’t be like me and trust the version string because it’s the one your scanner reports. Here at Firemind we’ve built the IT Operations Engine to patch your operating systems in the right order, following your rules, and to assess the repositories that build your images, with fixes arriving as reviewed pull requests. Higher-risk changes wait for a human, and every action is recorded. Scope a pilot on one AWS account or cluster, and see where NGINX runs across your cloud and the code that builds it, with every fix held for your approval.
Sources
- Wiz Research, “Unauthenticated Remote Code Execution in NGINX: CVE-2026-123633 and CVE-2026-123007”, Nir Ohfeld and Tomer Schwartz, 18 November 2026: five teams found the flaws independently; both need HTTP/2; CVE-2026-123007 dates to 1.9.5 (2015) and was shown executing code on NGINX serving static files; fixed versions; disabling HTTP/2 as the interim fix; applications exposed even when their own code is secure.
- Xint, “NGINX2Root: New Remote Code Execution Vulnerabilities in nginx Require Urgent Patching”, 18 November 2026: the nginx2root name; version strings and distro backports; the Ubuntu package comparison; the ALPN check; the chain with a Linux kernel SCTP flaw to escape a Docker container and get root on the host; the SCTP module block.
- CERT-EU Security Advisory 2026-022, “Critical vulnerabilities in Nginx”, 18 November 2026: five unauthenticated code-execution flaws; HTTP/2 on a listening port as the only precondition for CVE-2026-123007; update swiftly, internet-facing first; disabling HTTP/2 mitigates two of the five; no exploitation in the wild reported at the time.
- The Register, “AI found more critical bugs in another widely used technology running 225M websites”, Jessica Lyons, 18 November 2026: F5’s statement that it had found no evidence of active exploitation; the chain working on AWS, Azure and Google Cloud.
- Azure/aks-app-routing-operator issue #623, opened 18 November 2026: a customer report that the managed app-routing image ships NGINX 1.27.1 with HTTP/2 on by default.
- Firemind case study: one security rule applied and verified across 21 security groups in 13 regions in 15 minutes.