
Secure application code won't protect you when the vulnerable part is the server in front of it.
CVE-2026-123633 and CVE-2026-123007 are two critical memory-safety flaws in how NGINX handles HTTP/2. Both can give an unauthenticated attacker code execution on the NGINX server itself. Exploitation doesn't need a bug in the website or API behind it, so an application with clean code can still be exposed.
Like Java4Shell a day earlier, this is a flaw in infrastructure your code runs behind, not in a package your code imports. That makes the first question the same: where is NGINX actually running, and is HTTP/2 turned on?
On November 18, 2026, F5 released fixes for two critical NGINX HTTP/2 vulnerabilities. Five research teams found and reported them independently, and both are rated 9.2 under CVSS 4.0.
Both can crash an NGINX worker. Code execution is possible where ASLR is disabled or can be bypassed. Researchers report that each flaw can also be used to read memory first, which helps an attacker get around ASLR.
Both bugs are old, CVE-2026-123007 dates back to NGINX 1.9.5 (2015), and CVE-2026-123633 to 1.19.7 (2021).
As of November 18, there were no public reports of exploitation. Rapid7 has published a proof-of-concept exploit for CVE-2026-123007 that achieves code execution against NGINX in a lab container. Wiz and Xint have not released theirs.
A deployment is exposed when all three of these are true:
http2 on; or listen ... http2 in your configuration.Reverse proxies are the most exposed for CVE-2026-123633. CVE-2026-123007 also affects NGINX serving content directly.
On NGINX Plus, the two CVEs were fixed in different builds. 37.0.8.1 or 37.3.0.1 covers both. NGINX builds shipped by OS distributions, ingress controllers and other products follow their own schedules and version labels, so take the fixed build from that vendor's advisory. Versions here follow the coordinated disclosure published on November 18.
NGINX rarely shows up as a dependency of your application. It arrives as a base image, an OS package, an ingress or gateway layer, or a component bundled inside a product. A source-built SBOM for the service behind it may not list NGINX at all, and a version string alone won't tell you whether HTTP/2 is on.
Ownership is the second gap. A single NGINX deployment often routes traffic for several applications, and it may belong to a platform team rather than any of the app teams it serves. When nobody owns the front door, nobody gets the alert.
The third gap is time. One of these flaws sat in the code for over a decade before four research teams found it independently. AI-assisted research is making old memory-safety bugs easier to find, for defenders and attackers alike, which shortens the window between a fix and the first exploit.
nginx -V against the live binary, not just the version in a Dockerfile, and check each config for HTTP/2.http2 off; in affected server blocks (NGINX 1.25.1 or later) and remove the http2 parameter from any listen directives. Test compatibility, reload, confirm old worker processes have exited, and turn HTTP/2 back on after you upgrade.WINDOW_UPDATE frames.Those signals are warning signs, not proof of compromise, and a quiet log doesn't prove safety. A CDN or WAF in front may reduce exposure but isn't a reliable control. The patch is the fix.
With Arnica, you can find where NGINX is used across Dockerfiles, IaC and container images, and route each fix to the right owner.
Container image mapping and scanning ties the images you run back to the repositories that build them. That surfaces packages introduced in base images, which never appear in source-code SBOM artifacts, and shows which team owns each image.
Agentic Pentesting tests the live application the way an attacker reaches it. It can surface risk that no manifest lists and prove whether it is exploitable in your environment.
If NGINX runs anywhere in your environment, start by finding which images and deployments include an affected build with HTTP/2 enabled, then patch the ones an attacker can reach first.
What are CVE-2026-123633 and CVE-2026-123007?
They are two critical memory-safety vulnerabilities in how NGINX handles HTTP/2, fixed by F5 on November 18, 2026. CVE-2026-123633 is a use-after-free in HTTP/2 connection handling, and CVE-2026-123007 is an integer underflow in HTTP/2 flow control. Both can allow unauthenticated remote code execution on the NGINX server and are rated 9.2 under CVSS 4.0.
Which NGINX versions are affected?
NGINX Open Source from 1.9.5 (CVE-2026-123007) and 1.19.7 (CVE-2026-123633) up to the fixed releases, 1.30.7 stable and 1.31.9 mainline. NGINX Plus R33 and 37.0.0 onward are affected; 37.0.8.1 or 37.3.0.1 fixes both CVEs.
Am I affected if HTTP/2 is disabled?
No. Both vulnerabilities require HTTP/2 to be enabled on the affected NGINX server. Check for http2 on; or the http2 parameter on listen directives.
How do I fix the NGINX HTTP/2 vulnerabilities?
Upgrade NGINX Open Source to 1.30.7 or 1.31.9, or NGINX Plus to 37.0.8.1 or 37.3.0.1. Rebuild every image that bundles NGINX, redeploy, and confirm old worker processes have exited. NGINX shipped by other vendors follows their own fixed versions.
Is there a workaround if I can't upgrade right away?
Yes. Set http2 off; in affected server blocks on NGINX 1.25.1 or later, remove the http2 parameter from listen directives, test, and reload. Turn HTTP/2 back on after upgrading.
Does secure application code protect against these vulnerabilities?
No. The flaws are in NGINX itself, so exploitation doesn't depend on a bug in the website or API behind it. Researchers showed code execution on NGINX serving static files with no backend application.
Are the NGINX HTTP/2 vulnerabilities being exploited?
As of November 18, 2026, there were no public reports of exploitation. Rapid7 has published a lab proof-of-concept exploit for CVE-2026-123007; Wiz and Xint have not released theirs. Packaged products that ship the same NGINX build to many customers are likely early targets.
A clean dependency search won't tell you whether the proxy in front of your app is exploitable. Arnica Agentic Pentesting tests the live application, using code, cloud, runtime and business context to prove which risks are exploitable. Every finding comes back confirmed or debunked, with the steps the agent took.
Learn more about agentic pentesting
Address 92% of risks before they ever reach production.
Book a demo