Skip to main content
November 19, 2026

NGINX HTTP/2 RCE: Find Every Proxy That Speaks h2

NGINX HTTP/2 RCE: Find Every Proxy That Speaks h2

When a link like Wiz's NGINX HTTP/2 write-up lands in a team chat, people react fast. Usually they start guessing which big project gets hit next. Nobody can answer that one. The question you can answer is narrower: which of our nginx processes speak HTTP/2 right now, and which of those sit in front of something we cannot afford to lose?

This post is a checklist for that question.

What Was Disclosed

F5 published two HTTP/2 fixes on November 18, 2026, in nginx 1.31.9 and 1.30.7. Wiz Research and several other teams reported them independently.

CVE-2026-123007 CVE-2026-123633
Bug Flow control window bug (send window underflow, per Wiz), heap out of bounds write Use-after-free in HTTP/2 connection handling
Vulnerable nginx OSS 1.9.5 to 1.31.8 1.19.7 to 1.31.8
Fixed nginx OSS 1.30.7 stable, 1.31.9 mainline 1.30.7 stable, 1.31.9 mainline
NGINX Plus fixed 37.0.8.1, 37.3.0.1 37.0.7.1, 37.2.0.1
Extra condition None beyond HTTP/2; Wiz calls it exploitable in default configs with h2 on and demonstrated it over TLS Proxying, and a worker running low on worker_connections
nginx.org severity Major Medium
CVSS 4.0 (F5 CVE record) 9.2 9.2

The sources come from nginx.org security advisories, the CVE records for CVE-2026-123007 and CVE-2026-123633, and the Wiz post. For NGINX Plus, the CVE records list R33 as the first affected release. Both bugs are remote and pre-auth. Wiz says "Both vulnerabilities require HTTP/2 to be enabled on the affected NGINX server."

The sources disagree on severity. The nginx changelog says the flow control bug "might allow an attacker to cause a segmentation fault in a worker process, or might have potential other impact". The F5 CVE records say code execution is possible if ASLR is disabled or bypassed. Wiz says it built working lab exploits, is not releasing them, and that ASLR does not protect against these issues. CISA's SSVC entry in both records says exploitation "none". Read that as "not yet", not as "low priority".

The 1.31.9 release also fixes other CVEs, including an HTTP/3 bug (CVE-2026-121810) and a resolver bug (CVE-2026-123634). That is one more reason to upgrade instead of living on a workaround.

Which Builds and Configs Are Exposed

A vulnerable version number alone does not mean you are exposed. Three things must all be true:

  1. The binary has HTTP/2 compiled in. The HTTP/2 module "is not built by default" and needs --with-http_v2_module. Most distro and official packages include it, so expect yes.
  2. A server block enables it. On 1.25.1 and later, HTTP/2 is turned on by http2 on; in http or server context. Older configs use the http2 parameter on listen, which still works but is deprecated. A fleet that grew over years often has both.
  3. A client can reach that listener. This can be h2 over TLS, or cleartext h2c with prior knowledge.

CVE-2026-123633 also needs proxying while the worker runs short of connections, so a busy reverse proxy that logs worker_connections are not enough is the likely target. CVE-2026-123007 needs neither, so a small static site with h2 on is still in scope.

Confirm It on a Running Fleet

Run these on each host, or inside each image:

nginx -v                                     # version
nginx -V 2>&1 | grep -o with-http_v2_module  # h2 compiled in?
nginx -T 2>/dev/null | grep -nE '^\s*(http2\s+on|listen\b.*\bhttp2\b)'   # h2 enabled where?

nginx -T dumps the full config, includes and all. That matters, because http2 on; set once in the http block turns it on for every server that does not say off.

Then check from the network, because the config on disk is not always what the running master loaded:

curl -s -o /dev/null -w '%{http_version}\n' --http2 https://login.example.com/          # 2 = h2
curl -s -o /dev/null -w '%{http_version}\n' --http2-prior-knowledge http://10.0.4.12/  # 2 = h2c
openssl s_client -alpn h2 -connect 10.0.4.12:443 </dev/null 2>/dev/null | grep ALPN

Probe both the public name and the origin's private address. If a CDN or load balancer terminates HTTP/2, the public probe only shows what the edge speaks. Wiz notes such an edge may mitigate parts of the exploit, depending on the implementation. It does nothing for anyone who can reach the origin directly.

Three places get missed in most inventories:

  • Containers. Every image that ships nginx has its own build. Run nginx -v inside the image, not on the node. Patching hosts does not fix an image built last month.
  • Bundled and forked nginx. Ingress controllers, API gateways and OpenResty-based stacks carry their own nginx core. Their version string may not look like 1.30.x. Check which core they bundle and whether their vendor has shipped the fix. The CVE records name only NGINX Open Source and NGINX Plus.
  • Distro packages. On November 18, Ubuntu listed its supported releases as vulnerable with priority High. The Debian tracker showed bookworm and trixie as vulnerable. Distros backport fixes, so compare against the distro's fixed build when it appears, not against 1.30.7. We hit the same trap with RHEL packages in the Apache mod_http2 triage.

If You Cannot Patch Today

The workaround Wiz gives is to turn HTTP/2 off: set http2 off; in affected server blocks (1.25.1 and later), remove the http2 parameter from any listen directive, reload, and confirm the old workers have exited. Turn h2 back on after the upgrade.

nginx -t && nginx -s reload
ps -o pid,lstart,cmd -C nginx    # no worker should start before the reload

That last check matters. A reload starts new workers, but old ones keep serving their open connections until those close. Long-lived HTTP/2 connections are exactly the traffic you want gone. The same applies after an upgrade: a new package with old workers still running is not patched.

Before you push http2 off everywhere, check what depends on HTTP/2:

  • gRPC. gRPC runs on HTTP/2. A server block using grpc_pass stops working for its clients when h2 is off. Patch those first, or accept downtime knowingly.
  • Browsers negotiate with ALPN and fall back to HTTP/1.1. They get slower page loads, not errors.
  • Shared listeners. Remove http2 from every listen on a shared address:port, then re-run the curl probe to prove h2 is gone.

Put an end date on the workaround in the ticket, or HTTP/1.1 quietly becomes permanent.

Decision Table

h2 on Reachable from In front of Proxies upstream Action
Yes Internet Identity provider, SSO, payment, admin Any Patch today. If you cannot, turn h2 off today, except on gRPC vhosts
Yes Internet Other apps Yes Patch today; this shape is exposed to both CVEs
Yes Internet Static content No Patch today or turn h2 off; CVE-2026-123007 still applies
Yes Internal or VPN only Any Any Patch in this window, with an owner and date on the ticket
No Any Any Any Patch on the normal cadence; add a check that h2 stays off
Unknown Any Any Any Run the probes above before you decide anything
Patched package, old workers Any Any Any Reload or restart and confirm the old PIDs are gone

Why the Proxy in Front of Your Identity Server Comes First

Teams often rank exposure by the app behind the proxy. For these bugs, rank by what passes through the proxy. An nginx that terminates TLS for your login service handles passwords, session cookies, authorization codes and tokens in plaintext. Its workers hold the TLS private key so they can run handshakes. Code execution in that worker reads all of it. It does not have to touch the identity server itself.

A hardened app behind the proxy does not lower the priority, because the attacker never needs to reach it. Patch the nginx in front of SSO, OAuth and admin panels first, then shared edge clusters, then the rest. The same logic applied to the NGINX Rift rewrite bug earlier this year.

Where an Agent Helps

In Akmatori you can write a skill that runs the checks above and returns the decision table. The SSH connector is read-only by default. To let the agent run nginx -T or nginx -v, add them as allowed command patterns. An allowed command can never contain $, a backtick or a redirect, so the skill should run each check as a plain separate command and leave out redirects such as 2>&1 and 2>/dev/null. With an Alertmanager or Grafana connector, an alert about nginx worker crashes can open an incident. The agent then reads the skill if its description names nginx and worker crashes. A human still decides when to reload production nginx.

Conclusion

Both bugs need HTTP/2. So the job is to find every nginx that speaks h2, including images, forks and the h2c listener nobody remembers. Upgrade to 1.30.7 or 1.31.9, or your distro's fixed build. Where you cannot upgrade yet, turn h2 off, confirm the old workers are gone, and put an end date on the ticket. Start with the proxy in front of your identity provider.

If your team wants AI-assisted incident workflows on your own infrastructure, Akmatori helps SRE teams investigate alerts, scope exposure, and draft safe mitigations for edge proxies. Powered by Gcore for global infrastructure reliability.

Automate incident response and prevent on-call burnout with AI-driven agents!