Back Upwind NGINX CVE-2026-123007 and CVE-2026-123633: Four Web-Facing RCEs, One Patch, and ...
F5 shipped fixes on 18 November 2026 for two HTTP/2 remote code execution flaws in NGINX, CVE-2026-123007 and CVE-2026-123633. What we found: the serious one is real but expensive, 128 GB on a single connection to reach code execution, and the most exposed deployments are community ingress-nginx clusters, which have HTTP/2 on by default and will never get a patch.
On 18 November 2026, F5 published fixes for both bugs, each rated CVSS 4.0 9.2. CVE-2026-123007 has been in NGINX since HTTP/2 support arrived in version 1.9.5 in 2015. We read both fix diffs and Rapid7’s public proof of concept to work out what an attack actually takes.
CVE-2026-123007 is the serious one, and a public PoC exists. Rapid7 published a crash trigger and a full RCE chain on GitHub, tuned for Debian 13 amd64 builds of 1.31.6 and 1.31.8.
It is not a cheap attack. It needs 60 GB on one connection to crash a worker and 128 GB to execute code. That traffic is also its best detection signal.
CVE-2026-123633 is much harder to reach. It needs an upstream and an attacker holding 15 of every 16 connection slots. There is no public PoC, and nginx.org rates it medium.
Community ingress-nginx is the most exposed. HTTP/2 is on by default, the project is archived, and the Rancher and Chainguard forks are unpatched too.
Turning off HTTP/2 blocks both bugs, and a reload is enough. It is the right stopgap everywhere except gRPC frontends.
Affected versions: NGINX 1.9.5 to 1.31.8 (CVE-2026-123007) and 1.19.7 to 1.31.8 (CVE-2026-123633), NGINX Plus R33 and later, and every release of community ingress-nginx.
Fixed versions: NGINX 1.30.7 (stable) and 1.31.9 (mainline); NGINX Plus 37.0.8.1 and 37.3.0.1 (CVE-2026-123007), 37.0.7.1 and 37.2.0.1 (CVE-2026-123633); NGINX Gateway Fabric 2.9.0; NGINX Ingress Controller 5.8.0. NGINX Plus R33 to R36 and community ingress-nginx have no fix.
What are CVE-2026-123007 and CVE-2026-123633?
Two memory-safety bugs in how NGINX handles HTTP/2, each fixed in a few lines.
CVE-2026-123007, an integer underflow. When a stream is cancelled, NGINX refunds its queued bytes to the connection’s send window without checking the HTTP/2 maximum. Once the window is above that maximum, the bounds check wraps around and accepts every WINDOW_UPDATE . The window eventually turns negative and reaches a memory copy in the TLS send path, which writes out of bounds.
CVE-2026-123633, a use-after-free. A connection that has received only part of a HEADERS frame is marked reusable, so NGINX may close it when connection slots run short. It stays marked after a request starts on it. If that request opens an upstream connection while slots are low, NGINX can close and free the very connection it is still using. The bug dates to 1.19.7 in 2021.
How hard is exploitation?
CVE-2026-123007 is practical but loud. The attacker must push the window past 2⁶³, which takes 2³² WINDOW_UPDATE frames of 13 bytes each, 52 GiB on one connection. NGINX’s flood check adds overhead on top of that. The totals are 60 GB to crash a worker and 128 GB for RCE.
What else it needs, from the code and the PoC:
HTTP/2 over TLS. The out-of-bounds copy happens only in the TLS send path.
Any response body. The PoC uses a one-byte Range request on a static file. No backend is needed.
Any modern x86-64 CPU. Rapid7’s chain uses AVX-512, but AVX2 also works.
CVE-2026-123633 needs the server under pressure. The request must open an outbound connection, through any *_pass directive or a resolver, while the attacker holds at least 15/16 of worker_connections , 960 of 1,024 by default.
We have not tested musl (Alpine) or ARM64 builds, and nobody has published results for them.
Anyone running an affected NGINX build with HTTP/2 on and reachable. That includes OpenResty, Tengine, Angie and freenginx, whose source still holds the unpatched code.
HTTP/2 is off unless configured. The module is compiled into the official image, the Alpine variant and the Ubuntu package, but no shipped config enables it. nginx -T output for http2 on or http2 on a listen line.
Exposure then depends on what sits in front:
Direct internet or an L4 load balancer such as NLB: exposed.
An L7 load balancer or CDN that terminates HTTP/2 (ALB, CloudFront, Cloudflare): the frames are rebuilt, which likely breaks CVE-2026-123007. Treat it as reduced risk, not a fix.
Which versions are affected?
NGINX Open Source, CVE-2026-123007: 1.9.5 to 1.30.6 and 1.31.0 to 1.31.8. Fixed in 1.30.7 and 1.31.9.
NGINX Open Source, CVE-2026-123633: 1.19.7 to 1.30.6 and 1.31.0 to 1.31.8. Fixed in 1.30.7 and 1.31.9.
NGINX Plus, CVE-2026-123007: affected from R33 and from 37.0.0. Fixed in 37.0.8.1 and 37.3.0.1.
NGINX Plus, CVE-2026-123633: affected from R33 and from 37.0.0. Fixed in 37.0.7.1 and 37.2.0.1.
NGINX Plus R33 to R36: affected by both CVEs, with no fix.
NGINX Ingress Controller: fixed in 5.8.0, where HTTP/2 is off by default. The 2026-lts-r11 release still ships NGINX 1.31.0 and is affected.
NGINX Gateway Fabric: fixed in 2.9.0, where HTTP/2 is on by default.
Community kubernetes/ingress-nginx : every release is affected and none will be fixed. It bundles NGINX 1.27.1, enables HTTP/2 with use-http2: true by default, and is archived. The Rancher and Chainguard forks are unpatched too.
Docker images and distribution packages: the official Docker image has no 1.30.7 or 1.31.9 tag yet. Debian, Ubuntu, RHEL, Amazon Linux, SUSE, Alpine, Fedora and Bitnami have no fixed packages yet.
When distribution fixes land, they will be backports, so nginx -v will keep showing the old version. Compare the full package release with your distribution’s advisory.
Is the container a boundary?
No. Other researchers chained CVE-2026-123007 with a still-unpatched Linux kernel bug in SCTP, present since Linux 4.15. From an NGINX worker inside a Docker container, they reached root on the host, attacking an AWS server over its public IP. Where SCTP is reachable from the container, one RCE in an ingress pod can become control of the node.
How can exploitation be detected?
Watch volume, not the access log. The trigger frames are never logged as requests.
One TLS connection with tens of GB inbound and little outbound. This is CVE-2026-123007 in progress.
worker process ... exited on signal 11 (or 6) in the error log. A crash, from a failed attempt or the 60 GB DoS.
worker_connections are not enough, reusing connections in the error log. The precondition for CVE-2026-123633.
262,000 TLS 1.3 KeyUpdates and 16 MB HEADERS frames. These are specific to the public PoC, but strong when seen.
An NGINX worker starting any program. The clearest sign of success.
What should security teams do now?
Upgrade and restart. Move to 1.30.7, 1.31.9, Ingress Controller 5.8.0, Gateway Fabric 2.9.0 or the fixed Plus builds. A new binary needs a restart or a USR2 binary upgrade, not just a reload.
Until then, turn off HTTP/2. Set http2 off; and remove http2 from listen lines (on versions before 1.25.1, only the listen change applies). A reload is enough. gRPC frontends can’t do this, since grpc_pass needs HTTP/2, and they are in CVE-2026-123633’s scope because they have an upstream.
On community ingress-nginx, set use-http2: "false" in the controller ConfigMap now, then migrate. No patch will come.
Cap per-connection inbound bytes at the edge as a weaker extra layer. Don’t rely on keepalive_time , which HTTP/2 checks only when a new stream starts.
Block the SCTP module where you don’t use it by adding install sctp /bin/false to a file in /etc/modprobe.d/ .
Your risk here depends on which images and pods run which NGINX build, whether HTTP/2 is on, and what stands in front of them. Upwind’s sensor inventories container images, packages and Kubernetes workloads with their known CVEs, and maps internet exposure, so you can find the reachable ingress-nginx pods first.
Because the trigger stays out of the logs, the reliable signals are runtime ones. Upwind watches process execution and lineage, so an NGINX worker that crashes repeatedly or spawns a shell stands out.
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
