Threat advisory 9 min read

nginx HTTP/2 remote code execution (CVE-2026-123633 and CVE-2026-123007): what to do now

Two critical HTTP/2 bugs in nginx allow unauthenticated code execution. Which versions are affected, the fix, the workaround, and what WAFio can and cannot do.

A blue hexagon with a white letter N split by a glowing red crack, next to the words NGINX and HTTP/2 RCE, in a dark server room

On 18 November 2026 F5 released nginx 1.30.7 and 1.31.9. The headline fixes are two HTTP/2 bugs, CVE-2026-123633 and CVE-2026-123007, both rated critical (CVSS 4.0 score 9.2) and both able to lead to code execution on the server without a login. Wiz Research, Rapid7 and several other teams found them independently and published their analyses the same day.

The same releases also fix three more memory bugs outside HTTP/2, in the DNS resolver, in HTTP/3 and in the mail proxy. Turning HTTP/2 off covers the first two bugs only, so the upgrade is what closes all of them.

nginx often sits in front of several applications at once and ends TLS and HTTP/2 for all of them. A bug at that layer reaches every site behind it.

Where the bugs sit on the request pathVisitorsHTTPS + HTTP/2nginxends TLS and HTTP/2both bugs are hereHTTP/1.1WAFio WAF agentreads requests, never sees HTTP/2 framesYour applicationSame server: the WAFio host agent watchesprocesses, files and outbound connections
Both bugs are in the HTTP/2 layer of nginx, before any request reaches the WAF agent behind it.

Which versions are affected

Both HTTP/2 bugs need HTTP/2 to be enabled. nginx with HTTP/2 off is not exposed to them.

BugAffected nginx open sourceFixed inNGINX Plus fixed in
CVE-2026-123633, use-after-free1.19.7 to 1.30.6, 1.31.0 to 1.31.81.30.7, 1.31.937.0.7.1, 37.2.0.1
CVE-2026-123007, send window underflow1.9.5 (2015) to 1.30.6, 1.31.0 to 1.31.81.30.7, 1.31.937.0.8.1, 37.3.0.1

NGINX Plus R34 to R36 get no fix, according to Rapid7's summary of the F5 advisory, so those installs have to move to a 37.x build. Distribution packages follow on their own schedule: check the changelog of the package you run, not only the upstream version number.

CVE-2026-123633: a connection freed while still in use

When nginx runs short of connection slots, it closes connections it believes are idle. The bug is in how an HTTP/2 connection gets on and off that list:

How CVE-2026-123633 frees a connection that is still in use1The client opens an HTTP/2 connectionand leaves it idle. nginx lists it asa connection it may close to free a slot.2A new request starts on that connectionThe bug: nginx does not take it offthat list.3The request is proxied, so nginx needsa connection to the backend whileconnection slots are scarce.4nginx closes the listed connection tofree a slot: the one the request isstill using.5The request carries on with freedmemory: a use-after-free.
The use-after-free needs HTTP/2 and requests that nginx forwards to a backend.

Wiz notes that the path where nginx fails to build its GOAWAY frame, for example at its internal frame limit, is one way the connection memory gets freed while processing continues. The bug needs requests that nginx forwards to an upstream, so it mostly concerns reverse proxies.

CVE-2026-123007: a flow-control counter that wraps

HTTP/2 flow control limits how much data a sender may have in flight. nginx keeps that allowance for the whole connection in h2c->send_window, an unsigned size_t, and the protocol says it may never exceed 2^31 − 1. Rapid7 and Equixly describe the chain like this:

How CVE-2026-123007 turns a flow-control counter into an out-of-bounds copy1The client slows its reads and pushesthe windows close to the maximumsend_window ≈ 2^31 − 12A SETTINGS frame pushes one stream pastthe limit. nginx resets only that stream(RST_STREAM) and keeps the connection.3Cleanup refunds the queued data to theconnection with no upper-bound checksend_window > 2^31 − 14Each WINDOW_UPDATE check now wraps inunsigned math, so the counter keepsgrowingMAX − send_window → huge5The send path reads the counter as asigned off_t: a negative length, andmemcpy over TLS copies out of bounds.
The trigger needs only HTTP/2 over TLS; Wiz reproduced it with static files and no backend.

Two details explain why this is more than a crash. RFC 9113 says a stream window above the maximum is a connection error, but nginx only resets the stream, so the attacker keeps the connection and its broken counter. And the length check in ngx_ssl_send_chain() only trims sizes that are too large, so a negative size passes straight to ngx_memcpy().

Is anyone exploiting it?

At the time of writing there is no confirmed exploitation in the wild. Wiz built lab exploits for official 1.31.7 and 1.31.8 builds and will not publish them. Rapid7 did publish a proof of concept for CVE-2026-123007, against a Docker image of nginx 1.31.8 with HTTP/2 and TLS. By Rapid7's own account it has limits:

  • It depends on one x86-64 glibc memcpy variant that is picked on CPUs with AVX-512.
  • It sends about 128 GB of HTTP/2 traffic over one connection: roughly two minutes at 10 Gbit/s and seventeen at 1 Gbit/s. Below about 300 Mbit/s the default one-hour connection lifetime runs out first.
  • Code runs as the nginx worker user, not as root.

Those limits will not hold for long. Plan the upgrade in days, not weeks.

The other three bugs in the same release

CVEWhereNeedsStopgap until you upgrade
CVE-2026-123634DNS resolver, use-after-freeresolver, a proxy_pass target built from the request, and a subrequest such as mirrorPoint proxy_pass at a fixed upstream group
CVE-2026-121810HTTP/3 (QUIC), use-after-freeHTTP/3 with requests forwarded to an upstreamRemove quic listeners and set http3 off;
CVE-2026-113425Mail proxy auth_http, use-after-freeThe mail proxy with proxy onAllow only trusted addresses to the mail listeners

What to do today

  1. Find every nginx you run, including the ones inside containers, ingress controllers and appliance images. nginx -v shows the version; nginx -T | grep -nE 'http2|http3|quic|resolver|mirror|auth_http' shows which of the risky features each one uses.
  2. Upgrade to 1.30.7 or 1.31.9 (or your distribution's fixed package), then restart, not only reload, so no old worker keeps running the vulnerable code.
  3. If you cannot upgrade today, turn HTTP/2 off on the affected servers: on nginx 1.25.1 and later set http2 off; in each server block, and remove the http2 parameter from every listen line. Clients fall back to HTTP/1.1. Test your applications, reload, and check that the old workers have exited. Apply the stopgaps in the table above for the other three bugs.
  4. Turn HTTP/2 back on once you are on a fixed build.

Where WAFio fits, and where it does not

A web application firewall reads requests. These bugs live one layer below, in how nginx handles HTTP/2 frames, connections and DNS answers, before a request exists to inspect. In the usual setup nginx sits in front, ends TLS and HTTP/2, and passes plain requests to the WAFio WAF agent behind it, as in the first diagram. The agent never sees the frames, so no WAF rule fixes this. The fix is the upgrade above.

The WAFio WAF agent itself does not use nginx: it is written in Rust and speaks HTTP/1 to the front end, so these bugs do not apply to it. The nginx in front of it still needs the patch.

What WAFio adds is a way to see and stop what an attacker does after getting in:

  • Runtime security. An nginx worker serves files and proxies requests; it has no reason to start a shell. A host agent policy on execve events for the UID your workers run as (often www-data or nginx), matching /bin/sh, bash, curl or wget, can alert or kill the process. A second policy on outbound connect events from the same UID catches a reverse shell calling home.
  • Egress blocklist. The eBPF host firewall can block outbound traffic to addresses and ports your servers never need, which takes away the easy way to pull a second stage or send data out.
  • Signs of an attempt. The public exploit pushes around 128 GB through one HTTP/2 connection, which stands out in any traffic graph. Failed attempts often crash a worker, so lines like worker process ... exited on signal 11 in the nginx error log are worth an alert while you patch.

None of this replaces the upgrade. It shortens the time an intruder goes unnoticed if a server was reached before you patched. The runtime security docs show how to write such a policy, and the firewall docs cover the egress blocklist.

Sources