Public PoC Rapid7 released a full RCE chain for CVE-2026-123007. Fixed in NGINX 1.30.7 / 1.31.9. Community ingress-nginx will never be patched. Check your version →
Public RCE PoC ● Patch available ingress-nginx: no fix coming

NGINX HTTP/2.
Remote code execution.

Two memory-safety bugs in NGINX's HTTP/2 code. CVE-2026-123007 is real but expensive: about 128 GB on a single connection to run code. The most exposed are community ingress-nginx clusters, with HTTP/2 on by default and no patch ever coming.

9.2CVSS 4.0, both CVEs
~128 GBon one connection for RCE
2015bug present since 1.9.5
CVE-2026-123007 · CVE-2026-123633 · Updated Nov 18, 2026 · Upwind Research & MDR

Am I exposed?

Tell us what you run. We'll tell you which bug reaches you.
NGINX: nginx -v. Config: nginx -T | grep http2.
The two bugs

One practical, one hard

Both are memory-safety bugs in HTTP/2 handling, each fixed in a few lines. Both are rated CVSS 4.0 9.2, but they are not equally reachable.

CVE-2026-123007
Integer underflow · since 1.9.5 (2015) · public PoC

Practical, but loud

A cancelled stream refunds bytes to the send window without checking the HTTP/2 maximum. The bounds check wraps, the window goes negative, and a copy in the TLS send path writes out of bounds.

  • HTTP/2 over TLS
  • Any response body. The PoC uses a 1-byte Range request on a static file; no backend needed
  • Any modern x86-64 CPU (AVX-512 or AVX2)
CVE-2026-123633
Use-after-free · since 1.19.7 (2021) · no public PoC

Needs a server under pressure

A connection with a half-received HEADERS frame stays marked "reusable". If its request opens an upstream connection while slots are low, NGINX can free the connection it's still using. nginx.org rates it medium.

  • HTTP/2
  • An upstream: any *_pass directive or a resolver
  • Attacker holds ≥ 15/16 of worker_connections (~960 of 1,024)

What CVE-2026-123007 costs an attacker, on one connection

52 GiBunavoidable
60 GBworker crash
128 GBcode execution

2³² WINDOW_UPDATE frames of 13 bytes each, plus NGINX's flood-check overhead. That much traffic on one connection is also the best way to spot the attack. musl (Alpine) and ARM64 builds haven't been tested by anyone yet.

Fixed versions

Patch, then restart

A new binary needs a restart (or a USR2 binary upgrade). A config reload keeps the old code running.

ProductAffectedFixed inStatus
NGINX Open Source1.9.5 – 1.30.6
1.31.0 – 1.31.8
1.30.7 · 1.31.9Fix out
NGINX PlusR33+ and 37.0.0+123007: 37.0.8.1 · 37.3.0.1
123633: 37.0.7.1 · 37.2.0.1
R33–R36: no fix
F5 NGINX Ingress Controllerbefore 5.8.0, 2026-lts-r115.8.0Fix out · HTTP/2 off by default
F5 NGINX Gateway Fabricbefore 2.9.02.9.0Fix out · HTTP/2 on by default
Community ingress-nginxevery release (NGINX 1.27.1)—Archived · no fix
Docker image & distrosDebian, Ubuntu, RHEL, Amazon Linux, SUSE, Alpine, Fedora, BitnamiNo fixed tag or package yetWaiting
⚠️
nginx -v will lie. Distro fixes arrive as backports, so the version string stays old. Compare the full package release against your distro's advisory. Forks (OpenResty, Tengine, Angie, freenginx) still carry the unpatched code.
Find every image running a vulnerable NGINX build →

Who's actually exposed

HTTP/2 is off unless you turn it on. The module ships in the official image, Alpine and Ubuntu packages, but no default config enables it. What sits in front matters:

ExposedNGINX reachable directly from the internet, or through an L4 load balancer such as NLB.
Reduced risk, not a fixAn L7 load balancer or CDN that terminates HTTP/2 (ALB, CloudFront, Cloudflare) rebuilds the frames, which likely breaks CVE-2026-123007.
Is the container a boundary?

No. One RCE can become the node.

Researchers chained CVE-2026-123007 with a still-unpatched Linux SCTP kernel bug (present since Linux 4.15). From an NGINX worker in a Docker container, they reached root on the host of an AWS server, over its public IP.

1

Internet → ingress

~128 GB of HTTP/2 frames over one TLS connection. Nothing in access logs.

2

Code exec in worker

Runs as the NGINX worker inside the container.

3

SCTP kernel bug

Reachable wherever the container can open an SCTP socket.

4

Host root

Container escape to the node, and from there the cluster.

Community ingress-nginx

  • HTTP/2 on by default (use-http2: true).
  • Archived. Rancher and Chainguard forks are unpatched too.
  • Now: set use-http2: "false" in the controller ConfigMap, then migrate.
Not its first time: IngressNightmare →

Close the kernel step

  • Block SCTP where you don't use it: install sctp /bin/false in a file under /etc/modprobe.d/.
  • No kernel patch exists yet.
How Upwind spots container escapes →
Do this now

Five steps for security teams

01

Upgrade & restart

1.30.7, 1.31.9, Ingress Controller 5.8.0, Gateway Fabric 2.9.0 or fixed Plus. Restart, don't just reload.

02

Turn off HTTP/2

Blocks both bugs. A reload is enough.

http2 off; # and drop http2 from listen
03

ingress-nginx

No patch will come. Disable HTTP/2, then migrate.

use-http2: "false"
04

Cap at the edge

Limit inbound bytes per connection. A weaker extra layer, since the attack needs ~60–128 GB.

05

Block SCTP

Removes the kernel step of the container escape.

install sctp /bin/false
⚠️ Two trapsgRPC can't turn off HTTP/2. grpc_pass needs it, and those frontends have an upstream, so they're in scope for CVE-2026-123633 too. Patch them first.Don't rely on keepalive_time. HTTP/2 only checks it when a new stream starts, so it won't cut a long attack connection. On versions before 1.25.1, only removing http2 from listen works.
Detect attempts

Watch volume, not the access log

The trigger frames are never logged as requests. These are the signals that work, in the order an attack produces them. No team watching around the clock? Upwind MDR →

In progress

One TLS connection, tens of GB in, little out

CVE-2026-123007 being exploited. The cost of the attack is its biggest tell.

Precondition

Connection slots running out

What CVE-2026-123633 needs.

worker_connections are not enough, reusing connections
Failed attempt

Worker crashes

A failed exploit, or the 60 GB denial of service.

worker process ... exited on signal 11   (or 6)
PoC fingerprint

~262,000 TLS 1.3 KeyUpdates and 16 MB HEADERS frames

Specific to the public PoC, but a strong signal when you see it.

Success

An NGINX worker starts a program

The clearest sign the attack worked. Upwind runtime detection flags it with full process lineage.

www-data  sh -c id; uname -a; cat /etc/passwd
See Upwind cloud detection & response →
Timeline

How we got here

HTTP/2 support ships, and with it the CVE-2026-123007 bug.

CVE-2026-123633 is introduced.

F5 publishes fixes (1.30.7 / 1.31.9). Rapid7 releases a public crash trigger and RCE chain for CVE-2026-123007.

We'll update this page as Docker tags and distro packages ship, and if exploitation is seen in the wild.

Related research

More from Upwind

Upwind customers

Find reachable ingress-nginx first

Upwind inventories images, packages and Kubernetes workloads, maps which NGINX is reachable from the internet, and flags a worker that crashes repeatedly or spawns a shell.