← Blog
nginxHTTP/2RCEWeb Server Security

nginx HTTP/2 RCE: CVE-2026-123007 patched, PoC public

nginx HTTP/2 flow-control bug CVE-2026-123007 enables unauthenticated RCE. Patch to 1.31.9 or 1.30.7 now; a working PoC is public. Full remediation.

Zero Hunt Research··7 min read

Published by Zero Hunt, an autonomous AI red team on an on-premise appliance running private AI: automated penetration testing for networks and infrastructure, black-box or gray-box, with a human approving every step that matters.

nginx is the front door of a large share of the public internet, and on 18 November 2026 F5 patched an unauthenticated remote-code-execution bug in the part of it almost everyone turns on: HTTP/2. CVE-2026-123007 is an integer underflow in HTTP/2 flow control that lets a remote attacker, with no credentials, corrupt a worker process's memory — crashing it, and under the right conditions executing code in its context. Two research teams found it independently, and one of them, Rapid7, has published a working proof-of-concept. The fix is a point release; the hard part is knowing where HTTP/2 is actually on.

Developing story — first published 18:39 CET (17:39 UTC), November 18, 2026. Updated as the vendor and CISA publish more.

At a glance

CVE CVE-2026-123007
Product / affected versions NGINX Open Source 1.9.5–1.31.8 with HTTP/2 enabled; NGINX Plus R33–R36 · 37.0.0–37.0.7.1 · 37.1.0–37.2.0.1
Fixed in NGINX OSS 1.31.9 (mainline) · 1.30.7 (stable); NGINX Plus 37.3.0.1 · 37.0.8.1 — R34–R36 have no fix yet as of Nov 18, 2026
CVSS 9.2 CRITICAL (CVSS 4.0) · 8.1 HIGH (CVSS 3.1) — both F5 SIRT; vector CVSS:4.0/AV:N/AC:H/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H
Exploited in the wild No — CERT-EU reports none; researchers demonstrated RCE in a lab
CISA KEV Not listed (catalog 2026.11.17)
Public PoC Yes — Rapid7 published a PoC achieving RCE against nginx 1.31.8; Xint withheld the final exploit steps
Official advisory F5 K000164058; nginx.org security advisories

What happened: an HTTP/2 flow-control underflow

HTTP/2 keeps a per-connection and per-stream "send window" — how many bytes nginx is allowed to push to a client before waiting for a WINDOW_UPDATE. The bug is a bookkeeping error in that accounting. When a response is cancelled, nginx credits the unsent bytes back to the connection window without checking whether that pushes the counter past its 2³¹−1 ceiling. Once it is over the ceiling, a later WINDOW_UPDATE subtraction wraps around to an enormous positive value. A SETTINGS frame then quietly terminates one stream while leaving the connection alive — where RFC 9113 would require a connection error — and the oversized window turns a length calculation negative. That negative length reaches a memory copy as a gigantic unsigned count, and the copy writes past the buffer into a neighbouring heap object, which is what makes code execution possible.

That is the mechanism the discoverers, Xint (who found it preparing for ZeroDay Cloud 2026 and reported it through HackerOne on 16 October) and Rapid7's Stephen Fewer, walk through. It is classified CWE-191 (integer underflow), and it lives entirely in the data plane: the flaw is reachable by any client that can open an HTTP/2 connection, with no authentication and no configuration on the attacker's side beyond speaking the protocol.

Who is exposed

You are a candidate if nginx is terminating HTTP/2 — which, for any site serving HTTPS with http2 on; or listen ... http2, is the default most people never turned off. CVE-2026-123007 affects NGINX Open Source from 1.9.5 through 1.31.8 and NGINX Plus across the R33–R36 and 37.x lines. The same 18 November release fixes a batch of other nginx memory-safety bugs (HTTP/3, resolver, and mail-proxy use-after-frees among them), so the upgrade is worth doing regardless of which modules you run.

One caveat that trips teams up: the version banner is not the whole answer. Distributions backport security fixes without bumping the upstream version, so an Ubuntu or Debian package reading 1.18.0 may already carry the patch — or may not. Check your distro's advisory, not just nginx -v.

How hard is this actually to exploit

This is where the honest reading matters, and why the CNA scored it CVSS 4.0 9.2 but with attack complexity High. A crash — a denial of service against the worker — is the easy, reliable outcome. Full remote code execution is real but demanding: Rapid7's proof-of-concept drove roughly 119 GiB of HTTP/2 traffic over a single long-lived connection against a 64-bit linux/amd64 nginx 1.31.8 container, and depends on the specific glibc memory-copy routine selected on x86-64 CPUs. On a 1 Gbit/s link that is on the order of 17 minutes of sustained traffic, and nginx's default one-hour connection lifetime means links slower than about 300 Mbit/s cannot finish in time. Proxied HTTP/2 connections likely will not retransmit in a way that cooperates.

"Alternative RCE exploitation strategies may be viable, but none are currently known." — Rapid7

None of that is a reason to wait. The barrier to a crash is low, a public PoC shortens the distance for everyone, and "hard today" has a way of becoming "scripted next week." Treat it as a patch-now, with the RCE caveat guiding how you prioritise, not whether you act.

Remediation

1. Am I affected? Confirm the version and whether HTTP/2 is actually negotiating:

nginx -v                                   # build version (but see backport caveat)
grep -R -nE 'http2|listen .*http2' /etc/nginx/   # is HTTP/2 configured?
openssl s_client -connect HOST:443 -alpn h2 </dev/null 2>/dev/null | grep -i alpn
# "ALPN protocol: h2" means HTTP/2 is live on that listener

2. Patch — exact fixed versions. NGINX Open Source: 1.31.9 (mainline) or 1.30.7 (stable). NGINX Plus: 37.3.0.1 (from 37.1.0–37.2.0.1) or 37.0.8.1 (from 37.0.0–37.0.7.1). NGINX Plus R34–R36 have no fixed release yet — those deployments must disable HTTP/2 or move to the 37.x line. After upgrading, restart nginx or do a supported binary upgrade — a config reload does not replace the running executable, and old workers can linger as "shutting down."

3. Can't patch now? Disable HTTP/2. On nginx 1.25.1+, set http2 off; in each affected server block and remove http2 from listen. On 1.9.5–1.25.0, drop http2 from listen 443 ssl http2;. Reload, then confirm with the openssl s_client -alpn h2 check above that h2 no longer negotiates. This mitigates CVE-2026-123007 and the HTTP/2 use-after-free in the same batch — it does not cover the HTTP/3, resolver, or mail-proxy bugs, so it is a stopgap, not a fix.

4. Hunt for compromise. There are no published IOCs — this is behavioural hunting (ATT&CK T1190, with T1499 for the crash path and T1059 if code ran):

  • Error log: worker process ... exited on signal, signal 11, or segfault; and coredumpctl list nginx. A worker crash has innocent causes too — weigh it with the rest.
  • An auditd execve rule scoped to the nginx worker's UID. A worker spawning /bin/sh or /bin/bash is the clearest sign the crash became code execution.
  • Traffic: a single long-lived TLS connection where inbound bytes climb while outbound stays small, alongside many short connections from the same IP. WINDOW_UPDATE/RST_STREAM frames are not logged as requests, so access logs will look quiet: ss -tinp state established '( sport = :443 )'.

5. Eradicate and verify. If you have evidence of worker-context execution, treat that host as a foothold: rebuild or clean from trusted media and rotate any secret the worker could read — upstream credentials, keys in the environment or config. After patching, confirm the running build changed (nginx -v plus a full restart), and verify the crash is no longer reproducible rather than assuming the version string settles it.

What is not known yet

  • No in-the-wild exploitation reported (CERT-EU, as of disclosure), and NVD's SSVC currently records exploitation as none. Whether attackers weaponise it at scale given the traffic and architecture preconditions is open.
  • NVD has not finished analysis — the record is Received, scored by F5 SIRT; the NVD base score may land differently.
  • NGINX Plus R34–R36 have no patched release at disclosure; F5 has not stated when one ships.
  • Non-x86-64, non-glibc, or hardened builds: whether RCE (as opposed to a crash) is achievable outside the demonstrated target is unproven.
  • CISA KEV: not listed as of catalog 2026.11.17.

Where Zero Hunt fits

A clean point release and a compensating control exist, so the real work here is sequencing and proof: which of your nginx fleet is genuinely reachable, which is actually running HTTP/2, which is on the unfixed R-line, and — before you schedule an emergency change window — which hosts a version match even makes exploitable. Zero Hunt's AI Remediation Advisor ranks findings by real exploitability (CISA KEV, then proven reachable in the campaign, then CVSS, then EPSS) and hands back the exact fixed version with rollout and rollback notes plus the disable-HTTP/2 stopgap, in the spirit of risk-based patching under BOD 26-04 rather than patching a thousand banners blind.

What tells you a banner is more than a banner is proof. Because CVE-2026-123007's exploitability hinges on build, architecture, whether the connection is proxied, and whether HTTP/2 is truly on, a version scan over-reports. Zero Hunt's autonomous AI red team runs gray-box against your exact installed version, fingerprints the real configuration, and attempts to drive the code path in an ephemeral, hardened container that never touches the appliance host — on-premise, on private models that never ship your configuration or credentials to a third party, with a human in the loop for anything that matters. The same proof that demonstrates the issue re-verifies the fix once you have patched. For the HTTP/3 reverse-proxy side of nginx, see our earlier analysis of CVE-2026-42530; for the other end of the HTTP/2 stack, the Apache mod_http2 RCE.

Every vulnerability CISA lists as exploited, with federal due dates: CISA KEV tracker →

Patching one CVE at a time? A CTEM program validates which exposures are actually exploitable before the next advisory: what CTEM is →

Is this exploitable in your environment?

Zero Hunt answers that on your own network: an autonomous AI red team on an on-premise appliance, running on private AI, black-box or gray-box, with a human approving every step that matters. Proof of what is exploitable, the fix, and signed evidence — no data leaves your perimeter.