TRENDING
A person sweeping a metal detector across a wet sandy beach with a sea fort on the horizon
November 22, 2026
How to Write YARA Rules With YARA-X to Hunt Malicious Python .pth Startup Files and Measure False Positives Before You Trust Them
A peeling black-and-white metal door standing ajar, with a dark empty hall and bright windows beyond
November 22, 2026
AWS’s “Found a Zero-Day on Their Own” Post Never Names the Product; the Advisory Behind Its Link Shows a Server That Ships With No Login, a README Now Saying Keep It Off the Internet, and a Fix Merged October 19 That PyPI and Docker Hub Still Do Not Carry
A hand paging through index cards in an open wooden library card-catalog drawer labeled Agriculture, with a second drawer open below to the right
November 22, 2026
Microsoft Tells Customers Unauthorized Parties Read Admin Email Addresses and Tenant Metadata Through an Entra Connect Health Flaw Introduced July 23, but Gives No Tenant Count or Start Date, and Its Own Docs Send Tenant Security Notices to the Same Kinds of Contacts
Swimmers in separate lanes of a blue swimming pool, seen from above the water, with lane lines running across the picture
November 22, 2026
How to Size a PostgreSQL Connection Pool in Python: Find Where Throughput Stops Growing, Split max_connections Across Workers, and Stop Slow Handlers From Hoarding Connections
Workers in caps lean over a 1913 Ford assembly line, each assembling parts as they pass on the moving line
November 22, 2026
OpenAI’s Performance Factory Turns Regression Response Into a Task Queue for Agents, but Its Numbers Stop at 600 Root-Caused Cases in 1,700 Investigations and a 70% Grade on 100 Suggestions, and in a Toy World Each of Its Reviewer’s Two Counterfactual Questions Can Return a Confident Wrong Answer
22 Nov 2026
SXZ.io SXZ.io
  • Home
Search the Site
Popular Searches:
Technology Amazon AI
Recent Posts
The sun setting over the sea behind a pyramid-shaped stone monument at Cape Soya, the northernmost point of Japan
How to Retire an API Endpoint in FastAPI: Send Deprecation and Sunset Headers, Find the Clients Still Calling It, Schedule Brownouts, and Block an Early Removal With oasdiff
November 22, 2026
Brass sextant in an open wooden case inside a museum display, an instrument for measuring the angle to the horizon
MIT Technology Review’s “Most Misunderstood Graph” Update Reports METR’s Judgment Findings Mostly Fairly, but It Never Says Each Point Is a 50% Success Rate or Gives the New 3.6-Month Doubling Time and 2.1 to 4.9 Month Range, and It Credits OpenAI With a Comparison METR Built
November 22, 2026
Brass telescope on a wooden tripod inside a white observatory dome with the roof slit open to the sky
Rackspace Says Leaked Credentials Let an Intruder Act as Legitimate Users in Some Spot Customer Organizations and Reach Their Kubernetes Secrets, a Platform9 Co-Founder Names a Breached Observability System as the Root Cause, and the Status Page Says Only “System Maintenance Operations”
November 22, 2026
SXZ.io SXZ.io
  • Home

Categories

Articles 310 Posts
News 316 Posts
Learning Hub 282 Posts
Home/News/NGINX Fixes Five Flaws F5 Scores 9.2, and Turning Off HTTP/2, the Workaround Most Coverage Offers, Blocks Only Two; Rapid7’s Exploit Needs About 128 GB of Traffic, Xint Says Its Own Worked Against Target Servers on AWS, GCP and Azure, and Red Hat and Debian Stable Had No Fix at Our Check
News

NGINX Fixes Five Flaws F5 Scores 9.2, and Turning Off HTTP/2, the Workaround Most Coverage Offers, Blocks Only Two; Rapid7’s Exploit Needs About 128 GB of Traffic, Xint Says Its Own Worked Against Target Servers on AWS, GCP and Azure, and Red Hat and Debian Stable Had No Fix at Our Check

nginx 1.31.9 and 1.30.7 fix five flaws F5 scores 9.2; disabling HTTP/2 blocks two, public exploit claims differ, and Debian stable and Red Hat had no fix at our check.

November 19, 2026 19 Min Read
32

F5 and the NGINX project released nginx 1.31.9 (mainline) and 1.30.7 (stable) on November 18 with fixes for ten vulnerabilities. Wordfence’s PSA and the headline of Wiz’s write-up name two critical HTTP/2 flaws, CVE-2026-123007 and CVE-2026-123633. F5’s records score five at 9.2, and Wiz’s own table, CERT-EU and Xint list five that can lead to code execution, and turning off HTTP/2, the workaround most coverage leads with, blocks only two of them.

Table Of Content

  • What shipped on November 18
  • Five paths to code execution, and what turning off HTTP/2 covers
  • One set of flaws, several severity readings
  • What the public exploit work shows, and what it leaves open
  • Rapid7: a proof of concept with stated limits
  • Xint: cloud demonstrations and a kernel chain
  • Wiz and Metnir: lab results
  • Exploitation status
  • Who had a fix at about 10:15 UTC on November 19
  • F5’s own products
  • The disclosure clock
  • A triage script for the five paths
  • The script
  • The test cases and their results
  • Running it on a real nginx dump
  • What we would do first
  • Patch, then restart
  • If you cannot patch yet, match the workaround to the path
  • Check your front doors, and watch for crashes
  • What the sources leave open
  • Sources

Three more details are easy to miss. The five records carry a CVSS 4.0 score of 9.2 and a CVSS 3.1 score of 8.1, while nginx.org’s own advisory table labels the HTTP/2 use-after-free “medium”. The public exploit work is not one result: Rapid7’s proof of concept needs about 128 GB of HTTP/2 traffic and a specific glibc code path, while Xint reports that its CVE-2026-123007 exploit worked against target servers on AWS, GCP and Azure under normal traffic. And at about 10:15 UTC on November 19, Ubuntu and Alpine had published fixes while Debian stable and Red Hat had not.

This post is built from the primary records: nginx.org, F5’s advisories and CVE records, the nginx GitHub repository, the researchers’ own write-ups and the distribution trackers. Checks marked as ours ran on the morning of November 19 (UTC). No exploit code was fetched, run or linked. For earlier entries in the same run of AI-found flaws in foundational software, see our posts on PHP2Shell, Apache mod_http2 and Java4Shell.

What shipped on November 18

nginx.org’s news page says “nginx-1.30.7 stable and nginx-1.31.9 mainline versions have been released”, with ten security fixes. The annotated release tags in the nginx GitHub repository are stamped 15:24 UTC, and F5 published the CVE records at 15:23 UTC. All ten are fixed in both releases, except CVE-2026-120993, which only affects 1.31.4 through 1.31.8 and so has no stable-branch entry.

CVE What the record calls it F5 CVSS 4.0 / 3.1 nginx.org label
CVE-2026-123007 Memory corruption in HTTP/2 flow control (integer underflow, out-of-bounds write) 9.2 / 8.1 major
CVE-2026-123633 Use-after-free in HTTP/2 9.2 / 8.1 medium
CVE-2026-121810 Use-after-free in HTTP/3 9.2 / 8.1 medium
CVE-2026-123634 Use-after-free during response processing in the resolver 9.2 / 8.1 medium
CVE-2026-113425 Use-after-free in the mail proxy module 9.2 / 8.1 major
CVE-2026-118899 Infinite loop reading a client request body over TLS 1.3 8.7 / 7.5 major
CVE-2026-120993 NULL pointer dereference with PROXY protocol v2 (stream, dynamic certificates) 6.9 / 5.3 minor
CVE-2026-125283 Buffer overflow with alias and try_files (the record says worker crash) 6.9 / 5.3 minor
CVE-2026-120967 Client certificate verification bypass in a stream server 6.3 / 4.8 medium
CVE-2026-123640 Insufficient validation in stream ssl_preread 6.9 / 5.3 minor

Scores come from F5’s CVE records (and NVD, which at our check showed only those two F5 scores and the status “Received”). The labels come from the nginx.org advisory table, which at our read did not explain its scale. One small quirk for anyone searching that page: the entry for the HTTP/3 flaw is written “CVE-CVE-2026-121810”, and its link points to the same mistyped ID.

Five paths to code execution, and what turning off HTTP/2 covers

F5’s descriptions of all five carry an ASLR condition: four say attackers can execute code “on systems with Address Space Layout Randomization (ASLR) disabled or when the attacker can bypass ASLR”, and the fifth, CVE-2026-123634, words it as ASLR being disabled or bypassed. F5’s impact lines are hedged (“potentially execute code remotely”, “may allow a code execution”), and its vectors rate attack complexity High. The researchers’ claims are stronger: Xint writes that each path “leads to remote code execution on its own”, and says it withheld the final steps. Only CVE-2026-123007 has several independent public demonstrations; Wiz and Metnir describe lab code execution for CVE-2026-123633, and for the other three the code-execution claim in the sources we read is Xint’s.

The configurations the paths need differ, and Xint’s table is the clearest statement of which workaround covers which path. Its summary line is blunt: “Note that disabling HTTP/2 addresses only two of the five nginx vulnerabilities.”

CVE Configuration the path needs First affected release http2 off blocks it?
CVE-2026-123007 HTTP/2 enabled on a listener. Wiz demonstrated code execution against nginx serving ordinary static files. 1.9.5 Yes
CVE-2026-123633 HTTP/2 for clients, requests proxied to a backend, and a worker running low on worker_connections 1.19.7 Yes
CVE-2026-121810 HTTP/3 (QUIC) for clients and requests forwarded to a backend 1.25.0 No
CVE-2026-123634 A proxy whose upstream name contains a variable, a resolver, and an SSI or mirror subrequest that resolves the same name 0.6.18 in the record; 1.13.4 for Xint’s tested path No
CVE-2026-113425 Mail proxy with proxy on and auth_http; the client closes during the authentication lookup 0.5.33 in the record; 0.5.15 for Xint’s path No

The CVE-2026-123633 row is the one that matters most for typical web hosting. F5’s record says “proxy client requests to a backend”, and a WordPress server that hands PHP to PHP-FPM is forwarding requests to a backend in the ordinary sense. We checked the 1.30.7 source: the proxy, FastCGI, uwsgi, SCGI and gRPC modules each refer to the shared ngx_http_upstream_init, and Metnir’s AddressSanitizer trace for this bug runs through ngx_http_upstream_connect. So our reading, which F5 does not state, is that any of the five *_pass directives counts as forwarding.

The ASLR wording matters too. Wiz writes: “Typical memory safety mitigations such as ASLR do not protect against these issues”, because attackers can use the bugs for memory reads before code execution. Metnir says its HTTP/2 use-after-free exploit ran with ASLR on.

One set of flaws, several severity readings

Source What it says about the five
F5 CVE records and advisories CVSS 4.0 9.2 (Critical) and CVSS 3.1 8.1 (High); both vectors have attack complexity High
Wiz, Metnir, Xint, Wordfence “Critical”, which is the 4.0 label
CERT-EU advisory 2026-022 CVSS 8.1 for each of the five
nginx.org advisory table major for CVE-2026-123007 and CVE-2026-113425; medium for CVE-2026-123633, CVE-2026-121810 and CVE-2026-123634
Red Hat Critical for CVE-2026-123007, CVE-2026-123633 and CVE-2026-121810; Important for CVE-2026-123634; CVSS 3.1 8.1 listed on the first two. No Red Hat page for CVE-2026-113425 at our check.
Ubuntu Priority high
CISA An SSVC entry dated November 18: Exploitation none, Automatable no, Technical Impact total. None of the five was in the KEV catalog (version 2026.11.18, released 19:17 UTC)
FIRST EPSS No score yet for any of the five at our check

The 8.1 and 9.2 are one rating in two systems, not a disagreement. We recomputed both vectors with the open-source cvss Python library and got 8.1 for CVSS 3.1 and 9.2 for CVSS 4.0. The same flaws with attack complexity Low would score 9.8 and 9.3. The ASLR condition behind the High complexity rating is also why F5’s impact lines say “potentially”. A reader who sees 8.1 High in one place and 9.2 Critical in another is seeing the same assessment through two formulas.

The nginx.org labels are a different matter. They rank the HTTP/2 use-after-free and the HTTP/3 use-after-free below the HTTP/2 integer underflow although F5 scores all three identically. The page does not say how the labels are chosen, so they should not be read as a downgrade of the flaws, but a patch-priority list built from that column alone would put two of the five code-execution paths in the middle tier.

What the public exploit work shows, and what it leaves open

Rapid7: a proof of concept with stated limits

Stephen Fewer of Rapid7, credited on CVE-2026-123007, published a proof of concept against a linux/amd64 Docker image of nginx 1.31.8 with TLS and http2 on. Rapid7’s post lists the limits itself. Exploitation “requires a significant amount of network traffic to be transmitted, approximately 128 GB of HTTP/2 data”, which takes two and a half minutes in the demo, about two minutes on a 10 Gbit/s network and seventeen on 1 Gbit/s. Rapid7 says nginx’s default connection lifetime is one hour, so a network slower than about 300 Mbit/s cannot finish. A proxied HTTP/2 connection will likely not forward the attacker’s frames in a usable way, and the technique depends on a specific x86-64 glibc memcpy implementation chosen at runtime from the avx512f and avx512vl CPU features. Rapid7 adds: “Alternative RCE exploitation strategies may be viable, but none are currently known.” The repository README says the proof of concept “will likely fail against a vulnerable NGINX server which has multiple concurrent clients”.

Our arithmetic reproduces the transfer figures: 128 GB is 1,024 Gbit, which is 102 seconds at 10 Gbit/s and 1,024 seconds (about 17 minutes) at 1 Gbit/s, and 1,024 Gbit over 3,600 seconds is 284 Mbit/s.

Xint: cloud demonstrations and a kernel chain

Xint, credited on all five records, reports a different picture for CVE-2026-123007. Its post says “our exploit works under concurrent normal traffic and across servers with different hardware specifications”, and a table marks success on target servers on AWS, GCP and Azure, tested from inside the server, across a private network, and over a public HTTPS endpoint. Xint also says it chained the nginx bug with a separate flaw in the Linux kernel’s SCTP subsystem to reach root, including out of a Docker container. It is withholding the kernel details until a patch exists, names no CVE for the kernel flaw, and says the bug has been in the kernel since 2018. Its interim advice for hosts that do not use SCTP is to block module loading with install sctp /bin/false in a file under /etc/modprobe.d/, and it adds that this “only prevents loading through modprobe”: it does not unload a loaded module or disable SCTP compiled into the kernel.

The two accounts are compatible, since Rapid7 says other strategies may exist, but Xint’s results are its own, and it omitted the final exploitation steps. What a defender can say today is narrower than either headline: a public proof of concept exists with severe practical limits, and a vendor reports a more capable unpublished one.

Wiz and Metnir: lab results

Wiz says it demonstrated code execution for CVE-2026-123007 on nginx serving static files, with no backend application or third-party module, and used its Wiz Atlas harness to find both HTTP/2 flaws, with Gemini 4 Carbon and ChatGPT 5.6 Sol to validate the impact. Metnir, whose CVE-2026-123633 work it calls Drainbreak, describes command execution as the unprivileged nginx user on a lab VM, and writes that “the run in the screenshot took about ten minutes, nearly all of it spent on ASLR.” Four of the six groups credited on the two HTTP/2 records (Ikotas Labs, Metnir, Wiz and Xint) describe AI tools in their own accounts or credit lines, which fits The Register’s framing of this release as another AI-found batch.

Exploitation status

An F5 spokesperson told The Register that “we have found no evidence of active exploitation of any of these vulnerabilities in the wild”. CERT-EU wrote on November 18 that “no exploitation in the wild has been reported yet”, and CISA’s SSVC entry says Exploitation none. Public code is spreading regardless. GitHub’s repository search, when we ran it on the morning of November 19, returned 21 repositories for “CVE-2026-123007” and 10 for “CVE-2026-123633”. The first ten results for the first query were all created between 11:16 UTC on November 18 and 09:46 UTC on November 19, and they include scanners, lab setups and projects that describe themselves as exploits. We did not open, run or link any of them, so we cannot say which work. The Rapid7 repository was created at 11:16 UTC on November 18 and last pushed at 15:23 UTC, the same minute the CVE records were published.

Who had a fix at about 10:15 UTC on November 19

Source Status at our check
nginx.org 1.31.9 (mainline) and 1.30.7 (stable)
Ubuntu Released for 20.04 ESM (1.18.0-0ubuntu1.7+esm5), 22.04 (1.18.0-6ubuntu14.23), 24.04 (1.24.0-2ubuntu7.20) and 26.04 (1.28.3-2ubuntu1.13) for four of the five paths; the HTTP/3 flaw is marked “code not present” on the first three and released on 26.04
Debian Tracker lists bookworm (1.22.1-9+deb12u10), trixie (1.26.3-3+deb13u10) and forky (1.30.6-3) as vulnerable for all five; sid (1.30.6-4) fixed
Red Hat RHEL 8, 9 and 10 and Red Hat Hardened Images listed Affected for CVE-2026-123007 and CVE-2026-123633, with no errata yet; Red Hat’s mitigation entry is to disable HTTP/2
Alpine The secdb files for 3.22 and 3.23 list nine of the ten CVEs as fixed in nginx 1.28.3-r8 (CVE-2026-120993 does not affect 1.28); edge has 1.30.7-r0
Docker official image Last pushed: 1.31.9 at 17:19 UTC and 1.31.9-alpine at 19:53 UTC on November 18; 1.30.7-alpine at 19:52 UTC on November 18; the default 1.30.7 tag and stable at 01:50 and 01:52 UTC on November 19

Version numbers alone do not settle patch status. Ubuntu 22.04 reports nginx 1.18.0 both before and after its fix, and as Xint puts it, “a backport can fix a vulnerability without changing the version reported by nginx -v”. Compare the full package version with the distribution’s advisory.

F5’s own products

The F5 advisory for CVE-2026-123007 is a JavaScript page; we read it in a headless browser. Its product table lists fixed releases for NGINX Plus 37.0.8.1 and 37.3.0.1, NGINX Instance Manager 2.25.0, F5 WAF for NGINX 5.16.2 (and 5.14.4), F5 DoS for NGINX 4.13.0, NGINX Gateway Fabric 2.9.0 and NGINX Ingress Controller 2026-lts-r11 and 5.8.0. It lists vulnerable versions but no fixed release on several other branches: NGINX Plus R34 to R36, NGINX App Protect WAF 5.x and 4.x, NGINX App Protect DoS 4.x, NGINX Gateway Fabric 1.x and NGINX Ingress Controller 4.x and 3.x. A footnote says F5 evaluates only versions that have not reached End of Technical Support. BIG-IP, BIG-IQ, F5OS and BIG-IP Next for Kubernetes are listed as not vulnerable, and F5 Distributed Cloud and NGINX One Console are marked as still being researched. NGINX Plus fixed releases differ per CVE, so check the second advisory as well.

The disclosure clock

Date (UTC) Event Source
Aug 29 and Sep 12 Metnir’s pipeline finds the HTTP/2 use-after-free; Metnir reports it to F5 on September 12 Metnir, Wordfence
Sep 17 The fix for that bug is authored (commit 2df48a501d, credit “Albert @ Vexera AI”, the name Metnir used before) GitHub
Oct 11 to Oct 27 Xint discovers its vulnerabilities (Oct 11) and reports them through HackerOne (Oct 16); the reports reach F5’s nginx developers (Oct 27) Xint
Oct 21 Wiz reports CVE-2026-123633; F5 says another team had reported it first Wiz
Oct 29 The fix for the flow-control bug is authored (commit e27f5fa84f, credit Xint.io) GitHub
Nov 9 to Nov 10 The nginx team asks Xint to validate the patch; Xint finishes on Nov 10; F5 sends Wiz a patch for review Xint, Wiz
Nov 11 Rapid7 first contacts F5. The 1.31.9 and 1.30.7 release commits are dated 17:48 and 18:08. Files in nginx.org’s package repositories are stamped 20:04 to 21:17 (HTTP Last-Modified on the RPMs and indexes) Rapid7, GitHub, our requests
Nov 12 Rapid7 discloses CVE-2026-123007 to F5 Rapid7
Nov 16 and Nov 17 F5 confirms Rapid7’s report and sets November 18; Wiz reports the second HTTP/2 flaw on Nov 16 and F5 says on Nov 17 that it matches CVE-2026-123007 Rapid7, Wiz
Nov 18, 15:23 to 15:24 CVE records published; release tags created CVE Services, GitHub

Our arithmetic from those dates: 67 days from Metnir’s report to the release, 33 from Xint’s, 7 from Rapid7’s first contact, and 62 days between the first fix being authored and the release. The sources do not explain why a fix authored on September 17 waited until November 18. By our count of the nginx.org table, it lists 36 entries with 2026 CVE IDs against 2 with 2025 IDs and 7 with 2024 IDs, and 26 of the 2026 entries predate this release, which matches Xint’s statement that “by November 17 this year, it had recorded 26, not counting nginx2root”. The three most recent security releases on the news page are September 15, October 14 and November 18.

One timestamp we cannot interpret: the nginx.org package files and indexes show build stamps a week before the announcement. We could not see from outside whether they were reachable before November 18, the Internet Archive’s index showed no captures of the RPM listing between November 5 and 19, and neither F5 nor nginx.org says. If your servers pull from nginx.org’s repositories, check what your package manager installed rather than assuming a date.

A triage script for the five paths

This is our own tool, and it is a triage aid, not an exploit check. It reads configuration text, for example the output of nginx -T, and prints which of the five paths the configuration could reach, with line numbers. It cannot see distribution backports, include files missing from the dump, configurations generated at runtime by an Ingress Controller, or regex locations that contain braces, and it counts http2 on as on even if a more specific block turns it off. Target: nginx 1.9.5 through 1.31.8 configurations; tested on Python 3.13.14.

The script

# nginx_paths.py: triage aid for the NGINX 1.31.9 / 1.30.7 security release (2026-11-18).
# Reads an nginx configuration text (for example the output of `nginx -T`) and lists which of the five
# code-execution paths the configuration could reach. It cannot see distribution backports, includes that
# are missing from the dump, or regex locations that contain braces. "possible" means "look closer".
import re
import sys

PASS = ("proxy_pass", "fastcgi_pass", "uwsgi_pass", "scgi_pass", "grpc_pass")


def tokens(text):
    """Yield (kind, value, line); kind is 'w' for a word, or one of { } ;"""
    i, line, n = 0, 1, len(text)
    while i < n:
        c = text[i]
        if c == "\n":
            line += 1
            i += 1
        elif c.isspace():
            i += 1
        elif c == "#":
            while i < n and text[i] != "\n":
                i += 1
        elif c in "{};":
            yield c, c, line
            i += 1
        elif c in "\"'":
            j = i + 1
            while j < n and text[j] != c:
                j += 2 if text[j] == "\\" else 1
            yield "w", text[i + 1:j], line
            i = j + 1
        else:
            j = i
            while j < n and not text[j].isspace() and text[j] not in "{};":
                j += 1
            yield "w", text[i:j], line
            i = j


def directives(text):
    """Return (context tuple, name, args, line) for every directive."""
    stack, words, out = [], [], []
    for kind, value, line in tokens(text):
        if kind == "w":
            words.append(value)
        elif kind == ";":
            if words:
                out.append((tuple(stack), words[0], words[1:], line))
            words = []
        elif kind == "{":
            stack.append(words[0] if words else "?")
            words = []
        else:
            if stack:
                stack.pop()
            words = []
    return out


def analyze(text):
    d = directives(text)
    http = [x for x in d if "http" in x[0]]
    mail = [x for x in d if "mail" in x[0]]
    passes = [x for x in http if x[1] in PASS]
    h2 = [x for x in http if (x[1] == "listen" and "http2" in x[2]) or (x[1] == "http2" and x[2][:1] == ["on"])]
    quic = [x for x in http if x[1] == "listen" and "quic" in x[2]]
    auth = [x for x in mail if x[1] == "auth_http"]
    proxy_on = [x for x in mail if x[1] == "proxy" and x[2][:1] == ["on"]]
    resolver = [x for x in http if x[1] == "resolver"]
    var_pass = [x for x in http if x[1] == "proxy_pass" and "$" in " ".join(x[2])]
    mirror = [x for x in http if (x[1] == "mirror" and x[2][:1] != ["off"]) or (x[1] == "ssi" and x[2][:1] == ["on"])]
    at = lambda xs: ",".join(str(x[3]) for x in xs[:4]) or "-"
    return {
        "CVE-2026-123007 (HTTP/2 flow control)": ("possible" if h2 else "not seen", "http2 at line " + at(h2)),
        "CVE-2026-123633 (HTTP/2 use-after-free)": ("possible" if h2 and passes else "not seen", "http2 " + at(h2) + ", backend pass " + at(passes)),
        "CVE-2026-121810 (HTTP/3 use-after-free)": ("possible" if quic and passes else "not seen", "quic " + at(quic) + ", backend pass " + at(passes)),
        "CVE-2026-123634 (resolver use-after-free)": ("possible" if resolver and var_pass and mirror else "not seen", "resolver " + at(resolver) + ", variable proxy_pass " + at(var_pass) + ", mirror or ssi " + at(mirror)),
        "CVE-2026-113425 (mail proxy use-after-free)": ("possible" if auth and proxy_on else "not seen", "auth_http " + at(auth) + ", proxy on " + at(proxy_on)),
    }


def fixed(version):
    m = re.match(r"(\d+)\.(\d+)\.(\d+)", version or "")
    if not m:
        return "unknown"
    major, minor, patch = map(int, m.groups())
    if (major, minor) == (1, 30):
        return "fixed" if patch >= 7 else "vulnerable"
    if (major, minor) == (1, 31):
        return "fixed" if patch >= 9 else "vulnerable"
    return "check the distribution advisory"


if __name__ == "__main__":
    text = sys.stdin.read() if sys.argv[1] == "-" else open(sys.argv[1], encoding="utf-8", errors="replace").read()
    for name, (state, why) in analyze(text).items():
        print(f"{state:9} {name}   [{why}]")
    if len(sys.argv) > 2:
        print("version", sys.argv[2], "->", fixed(sys.argv[2]))

The tokenizer skips comments and keeps quoted strings whole, which matters because http2 can appear in a comment or a header value. It tracks block nesting so that listen 8443 ssl http2 inside stream is not flagged. The version helper only knows the two fixed lines (1.30.7 and 1.31.9); anything else, such as a distribution’s 1.24.0, returns a reminder to read the distribution advisory.

The test cases and their results

# run_cases.py: eight made-up configurations for nginx_paths.py, with the result each one must give.
import sys
from nginx_paths import analyze, fixed

CASES = {
    "A new-style http2 plus proxy": (
        "http { server { listen 443 ssl; http2 on; location / { proxy_pass http://app; } } }",
        {"123007": True, "123633": True, "121810": False, "123634": False, "113425": False}),
    "B legacy listen http2, static only": (
        "http { server { listen 443 ssl http2; root /srv/www; } }",
        {"123007": True, "123633": False, "121810": False, "123634": False, "113425": False}),
    "C http2 off, php-fpm": (
        "http { server { listen 443 ssl; http2 off; location ~ \\.php$ { fastcgi_pass unix:/run/php.sock; } } }",
        {"123007": False, "123633": False, "121810": False, "123634": False, "113425": False}),
    "D http3 plus proxy": (
        "http { server { listen 443 quic reuseport; listen 443 ssl; location / { proxy_pass http://app; } } }",
        {"123007": False, "123633": False, "121810": True, "123634": False, "113425": False}),
    "E resolver, variable proxy_pass, mirror": (
        "http { resolver 10.0.0.2; server { location / { mirror /m; proxy_pass http://$host; } location = /m { internal; proxy_pass http://$host; } } }",
        {"123007": False, "123633": False, "121810": False, "123634": True, "113425": False}),
    "F mail proxy with auth_http": (
        "mail { auth_http 127.0.0.1:9000/auth; server { listen 993 ssl; protocol imap; proxy on; } }",
        {"123007": False, "123633": False, "121810": False, "123634": False, "113425": True}),
    "G http2 only in a comment and a string": (
        "http { server { listen 443 ssl; # http2 on;\n add_header X-Note 'listen 443 http2;'; root /srv/www; } }",
        {"123007": False, "123633": False, "121810": False, "123634": False, "113425": False}),
    "H http2 inside stream, not http": (
        "stream { server { listen 8443 ssl http2; proxy_pass backend:9000; } }",
        {"123007": False, "123633": False, "121810": False, "123634": False, "113425": False}),
}

bad = 0
for name, (conf, want) in CASES.items():
    got = {k.split()[0][9:]: v[0] == "possible" for k, v in analyze(conf).items()}
    flags = " ".join(k for k, v in got.items() if v) or "none"
    ok = got == want
    bad += not ok
    print(f"{'ok  ' if ok else 'FAIL'} {name:42} -> {flags}")
for v in ["1.30.6", "1.30.7", "1.31.8", "1.31.9", "1.24.0"]:
    print("version", v, "->", fixed(v))
print("all cases matched" if not bad else f"{bad} case(s) failed")
sys.exit(1 if bad else 0)

Run python3 run_cases.py in the same folder:

ok   A new-style http2 plus proxy               -> 123007 123633
ok   B legacy listen http2, static only         -> 123007
ok   C http2 off, php-fpm                       -> none
ok   D http3 plus proxy                         -> 121810
ok   E resolver, variable proxy_pass, mirror    -> 123634
ok   F mail proxy with auth_http                -> 113425
ok   G http2 only in a comment and a string     -> none
ok   H http2 inside stream, not http            -> none
version 1.30.6 -> vulnerable
version 1.30.7 -> fixed
version 1.31.8 -> vulnerable
version 1.31.9 -> fixed
version 1.24.0 -> check the distribution advisory
all cases matched

Each case states which paths must be reported and the runner exits with a failure if any answer differs. Cases G and H are the traps: a directive named in a comment or a string, and http2 in a different protocol block.

Running it on a real nginx dump

We also ran it on output from a real nginx. This is the Windows build of nginx 1.30.6, which is one of the affected versions, with this configuration:

worker_processes 1;
events { worker_connections 64; }
http {
    server {
        listen 127.0.0.1:8081;
        http2 on;
        location / { proxy_pass http://127.0.0.1:9; }
    }
}

Dropping standard error keeps the “syntax is ok” lines out of the text, so the command is the one below (we pointed the Windows build at the file with its -p and -c options):

nginx -T 2>/dev/null | python3 nginx_paths.py - 1.30.6
possible  CVE-2026-123007 (HTTP/2 flow control)   [http2 at line 7]
possible  CVE-2026-123633 (HTTP/2 use-after-free)   [http2 7, backend pass 8]
not seen  CVE-2026-121810 (HTTP/3 use-after-free)   [quic -, backend pass 8]
not seen  CVE-2026-123634 (resolver use-after-free)   [resolver -, variable proxy_pass -, mirror or ssi -]
not seen  CVE-2026-113425 (mail proxy use-after-free)   [auth_http -, proxy on -]
version 1.30.6 -> vulnerable

The line numbers refer to the dump, whose first line is the # configuration file comment, so configuration line 6 shows as 7. The result says exactly what the configuration implies: HTTP/2 on and a backend pass, so the first two paths are possible, and nothing for HTTP/3, the resolver or the mail proxy.

What we would do first

This section is our reading, not F5’s guidance.

Patch, then restart

Install nginx 1.30.7 or 1.31.9, or your distribution’s fixed package, and recreate containers from patched images. Xint warns: “A configuration reload alone does not replace the running executable.” nginx’s control documentation lists HUP as starting new workers with a new configuration and USR2 as upgrading the executable, so use a restart or the documented binary upgrade. Wordfence’s point for WordPress owners holds: “Updating WordPress, plugins, themes, or Wordfence will not patch NGINX”, and “your host or server administrator must install the server fix”.

If you cannot patch yet, match the workaround to the path

For CVE-2026-123007 and CVE-2026-123633, on nginx 1.25.1 or later, set http2 off; in each affected server block and remove any http2 parameter from listen lines (the directive defaults to off and appeared in 1.25.1); on older versions remove http2 from listen:

server {
    listen 443 ssl;
    http2 off;   # nginx 1.25.1 or later; remove "http2" from listen lines too
}

Then confirm HTTP/2 is gone. Xint’s check prints nothing once HTTP/2 is off, and also nothing if the connection fails, so run it against a control too. Run against sxz.io, which serves HTTP/2, it prints:

openssl s_client -connect HOST:443 -servername HOST -alpn h2 </dev/null 2>&1 | grep -F 'ALPN protocol: h2'
ALPN protocol: h2

That does not touch the other three paths. Xint’s mitigations are to remove quic listeners, set http3 off and clear any HTTP/3 Alt-Svc header for CVE-2026-121810; to point proxy_pass at a fixed upstream group instead of a variable for CVE-2026-123634 (it notes that turning off mirror alone is not enough, since SSI subrequests may reach the same bug); and, for the mail proxy, to keep untrusted clients away by binding the mail listeners to a loopback address or allowing only trusted IP addresses at the firewall. Turning HTTP/2 off also breaks native gRPC, which needs it, and a reload applies the change only to new connections: Xint notes that workers still serving open HTTP/2 connections show as “nginx: worker process is shutting down” and that a restart is needed if they linger.

Check your front doors, and watch for crashes

A public HTTP/2 connection may end at a CDN while an origin nginx is reachable directly, and Wordfence advises checking the origin and any direct route to it. CERT-EU suggests watching the error log for “worker process … exited on signal”, with the caveat that it is “not a high-fidelity indicator of successful remote code execution but rather an indication of exploitation attempts”. Xint adds that an attack on CVE-2026-123007 may leave little in the access logs, because the WINDOW_UPDATE and RST_STREAM control frames it uses are not logged as requests.

What the sources leave open

  • Whether anyone is exploiting these flaws. F5, CERT-EU and CISA’s SSVC entry say no evidence yet, and public code is circulating.
  • Which of the public exploit results hold beyond the researchers’ own setups, and whether Xint’s kernel chain has a fix or a CVE.
  • When Red Hat and Debian stable will ship fixes, and what fix F5 will offer on the branches where its table lists none.
  • Whether F5 Distributed Cloud and NGINX One Console are affected; F5 says it is still researching.
  • Why the headlines carry two CVEs when F5’s own scoring puts five at 9.2.

Sources

  • nginx.org: security advisories, news, HTTP/2 module, control documentation
  • F5: K000163926 and K000163928; CVE records CVE-2026-123007, CVE-2026-123633, CVE-2026-121810, CVE-2026-123634 and CVE-2026-113425
  • Researchers: Wiz, Metnir, Xint, Rapid7
  • Coverage and advisories: Wordfence, CERT-EU, The Register
  • Distribution status: Ubuntu, Debian, Red Hat, Alpine secdb, Docker Hub
  • Other: CISA KEV, NVD, nginx commits 2df48a501d and e27f5fa84f

Tags:

CVE-2026-123007F5HTTP/2NGINXRemote Code ExecutionUse-After-FreeWiz

Share

A clear plastic reel of brown magnetic recording tape on a black background, shown face on with its three spokes
Previous Post

How to Recover PostgreSQL 18 to the Moment Before a Bad DELETE: WAL Archiving, pg_basebackup, Point-in-Time Recovery, and a Restore Drill You Can Test

Jagged granite ridgeline of Sawtooth Ridge above a fenced pasture with grazing cattle
Next Post

OpenAI’s Second Research-Acceleration Report Turns an 11x Rise in Agent Time Into a “Jagged Frontier,” but Its +62% Spend Jump Spans 18 Days From a Baseline Still Holding Part of a Harness Dip, Four of Seven Success Buckets Have Overlapping Bands, and Experiments per Experimenter Rose 1.34x Since May

No Comment! Be the first one.

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Latest
22 Nov
How to Retire an API Endpoint in FastAPI: Send Deprecation and Sunset Headers, Find the Clients Still Calling It, Schedule Brownouts, and Block an Early Removal With oasdiff
22 Nov
MIT Technology Review’s “Most Misunderstood Graph” Update Reports METR’s Judgment Findings Mostly Fairly, but It Never Says Each Point Is a 50% Success Rate or Gives the New 3.6-Month Doubling Time and 2.1 to 4.9 Month Range, and It Credits OpenAI With a Comparison METR Built
Trending
November 22, 2026
How to Retire an API Endpoint in FastAPI: Send Deprecation and Sunset Headers, Find the Clients Still Calling It, Schedule Brownouts, and Block an Early Removal With oasdiff
November 22, 2026
MIT Technology Review’s “Most Misunderstood Graph” Update Reports METR’s Judgment Findings Mostly Fairly, but It Never Says Each Point Is a 50% Success Rate or Gives the New 3.6-Month Doubling Time and 2.1 to 4.9 Month Range, and It Credits OpenAI With a Comparison METR Built
November 22, 2026
Rackspace Says Leaked Credentials Let an Intruder Act as Legitimate Users in Some Spot Customer Organizations and Reach Their Kubernetes Secrets, a Platform9 Co-Founder Names a Breached Observability System as the Root Cause, and the Status Page Says Only “System Maintenance Operations”
November 21, 2026
How to Write YARA Rules With YARA-X to Hunt Malicious Python .pth Startup Files and Measure False Positives Before You Trust Them
November 21, 2026
AWS’s “Found a Zero-Day on Their Own” Post Never Names the Product; the Advisory Behind Its Link Shows a Server That Ships With No Login, a README Now Saying Keep It Off the Internet, and a Fix Merged October 19 That PyPI and Docker Hub Still Do Not Carry
November 21, 2026
Microsoft Tells Customers Unauthorized Parties Read Admin Email Addresses and Tenant Metadata Through an Entra Connect Health Flaw Introduced July 23, but Gives No Tenant Count or Start Date, and Its Own Docs Send Tenant Security Notices to the Same Kinds of Contacts

Related Posts

Rows of server racks in a data center representing network infrastructure targeted by botnets
News

C0XMO Botnet Shows Why Old Router Firmware Still Matters

June 7, 2026
Close-up of a USB flash drive, representing physical data-theft risk in office security incidents
News

Fake IT Support Is Now Walking Through the Front Door

June 7, 2026
A phone security app on a smartphone resting on a laptop keyboard.
News

Everest Forms Pro Flaw Is Being Exploited to Create Rogue WordPress Admins

June 7, 2026
A phone secured by a padlock, illustrating AI data-leak containment and security controls.
News

OpenAI’s Lockdown Mode Is a Data-Leak Brake, Not a Prompt-Injection Cure

June 8, 2026
SXZ.io SXZ.io
  • [email protected]

Categories

Articles
Learning Hub
News

All Rights Reserved by SXZ.io ©2026