NGINX2Root: Xint says five flaws give code execution, Wiz two, Rapid7 one at 128 GB; Censys saw none
Four research teams published accounts of the NGINX flaws fixed on 18 November 2026: Xint says five give code execution, Wiz two, Rapid7 built one proof of concept needing about 128 GB of traffic, and Censys saw none in its tests. This site read the saved pages on 19 November.
By Parminder Kumar Sharma · LinkedIn · · 27 min read
Four accounts of one release count the code execution flaws as five, two, one and none
Xint's page says five of the NGINX flaws fixed on 18 November 2026 each "leads to remote code execution on its own". Wiz's summary names two of them, CVE-2026-123633 and CVE-2026-123007, as critical unauthenticated remote code execution flaws. Rapid7's page describes a proof of concept for one, CVE-2026-123007, and Censys's test post, which covers all five, lists no code execution for any of them. Counted from the four pages, that is five, two, one and none (derived). The same five flaws carry CVSS 3.1 8.1 and CVSS 4.0 9.2 in F5's CVE records, and nginx.org calls three of them medium and two major, as brief 742 reported from its saved copies.
What that does not establish. None of this says how many servers can be taken over today. Rapid7's page says its method needs about 128 GB of HTTP/2 data, will probably fail through a proxy and depends on one CPU feature set. Censys's small test reached a corrupted state for one flaw and crashed nginx on others, but produced no code execution. The technique sections of the Xint, Wiz and Rapid7 pages are removed from the copies this site saved, so this site cannot check any of the code execution claims, ranks none of the four accounts and picks no rating.
A sentence in brief 742 that has to be revisited. Brief 742 said of ASLR that this site "does not claim the flaw can be exploited on a default build with ASLR on, and no saved page claims it either". That described the nginx.org and F5 pages saved at about 08:56 UTC on 19 November. The pages saved at about 12:15 UTC are the ones that now speak to it. Wiz's page says its 21 October report of CVE-2026-123633 included "confirmed remote code execution with an ASLR bypass" and that typical mitigations such as ASLR "do not protect against these issues". F5's records still say code can run where ASLR is disabled or the attacker can bypass it. Rapid7's kept text names a CPU and library dependency instead. So the sentence was true of the pages 742 had and is not true of the pages saved since.
Whose words these are. F5 owns NGINX and sells NGINX Plus, BIG-IP and WAF products; Xint, Wiz, Rapid7 and Censys are security companies. Xint's page closes by inviting readers to have it scan their own code, Wiz's names its own scanning harness, and Censys's links its own tracker and advisory; the saved Rapid7 text carries no sales line. Each is a researcher or vendor describing its own work, and this site reports each claim as that page's own. This site names no individual and no victim: none is named as hit.
What each page says it did
The table sets down, in each page's own terms, what the researchers say they did and what they say about code execution. Nothing in it has been repeated by this site.
What each saved page says it did and says about code execution, read on 19 November 2026. Short quotations are exact.
| Page | What it says it did | What it says about code execution |
|---|---|---|
| Xint (18 November, written before the release) | Found five flaws; a Xint researcher found CVE-2026-123007 while preparing for ZeroDay Cloud 2026, and "a fully automated Xint scan" found the other four; chained the first with a separate Linux kernel flaw. | Each of the five leads to remote code execution on its own; the chain gave root on the host, also from a Docker container, in its tests (the page's video was not seen). |
| Wiz | Found two flaws with its Wiz Atlas harness; its researchers used Gemini 4 Carbon and ChatGPT 5.6 Sol to validate impact (names as printed). | Two critical unauthenticated flaws, rated 9.2 (CVSS 4.0); code execution demonstrated on NGINX serving static files for CVE-2026-123007; ASLR "do not protect". |
| Rapid7 | Built a proof of concept for CVE-2026-123007 against nginx 1.31.8 in a linux/amd64 Docker image with http2 on and TLS. | Code execution as the worker, needing about 128 GB of HTTP/2 data, a direct connection and one x86-64 glibc memcpy path; "none are currently known" for other strategies. |
| Censys (test post) | Tested all five in a lab; made a compact cleartext HTTP/2 test for CVE-2026-123007, the whole capture 5.69 MB; reproduced use-after-free conditions for three of the other four. | None of the reproductions resulted in remote code execution; it has not verified the kernel flaw or the chain. |
| F5 (index page, 18 November) | Lists 22 articles (derived count: 10 NGINX, 12 BIG-IP) as an index; the per-CVE articles were not fetched. | Nothing on the index page; the code execution wording is in the CVE records 742 read. |
| SecurityWeek (secondary, 19 November) | Reports the four accounts and F5's patches. | Relays them; it also says F5 "makes no mention" of exploitation in the wild. |
Two of the discovery claims are the researchers' own statements and this site did not check them: Xint says a fully automated scan found four of its five without human guidance, and Wiz says its harness found both of its flaws and that the two named models confirmed that each could let an unauthenticated attacker run code. Wiz also says Wiz Research, Ikotas Labs, Metnir, Rapid7 and Xint "independently uncovered" its two flaws, so the number of teams that found each bug depends on which page is read: Wiz says five teams for two flaws, and Xint writes only about itself.
The calendar: 33 days for one report, 7 for another, one release
Xint's disclosure timeline puts its report to F5 on 16 October and the scheduled patch release on 18 November: 33 days (derived: 15 days left in October plus 18). Rapid7's first outreach to F5 was on 11 November, 7 days before the release (derived). Wiz reported the first of its two flaws on 21 October, 28 days before the release, and the second on 16 November, 2 days before it (derived). The Xint span is 4.7 times the Rapid7 span (derived: 33 divided by 7).
Intervals between the dates the saved pages print (derived by this site; every date is as printed, in 2026).
| Interval | Dates | Days (derived) |
|---|---|---|
| Xint, found to reported | 11 Oct to 16 Oct | 5 |
| Xint, reported to HackerOne validation done | 16 Oct to 20 Oct | 4 |
| Xint, validation done to delivered to F5's nginx team | 20 Oct to 27 Oct | 7 |
| Xint, delivered to release | 27 Oct to 18 Nov | 22 |
| Xint, patch validation done to release | 10 Nov to 18 Nov | 8 |
| Xint, reported to release | 16 Oct to 18 Nov | 33 |
| Xint, found to release | 11 Oct to 18 Nov | 38 |
| Wiz, first report to release | 21 Oct to 18 Nov | 28 |
| Wiz, first report to F5's patch supplied for review | 21 Oct to 10 Nov | 20 |
| Wiz, second report to release | 16 Nov to 18 Nov | 2 |
| Rapid7, first outreach to F5 confirming validation | 11 Nov to 16 Nov | 5 |
| Rapid7, first outreach to release | 11 Nov to 18 Nov | 7 |
| Xint's report to Wiz's report of the second flaw | 16 Oct to 16 Nov | 31 |
What the intervals do not show. The pages print different kinds of dates: Xint's are when it found and reported, Wiz's and Rapid7's are when they reported, and none of the three says which flaws each date covers beyond what is named. For CVE-2026-123633 Wiz's 21 October entry says F5 confirms that "another research team had previously reported the same vulnerability"; the page does not say which team. Xint's 16 October date precedes it, but this site cannot say from the pages that Xint is the team meant. Ikotas Labs and Metnir are named by Wiz as independent finders and print no dates on any saved page. The 10 November entries are both on the record: Wiz's page says F5 supplied a patch for CVE-2026-123633 for review that day, and Xint's says its patch validation finished that day. The pages do not say whether it was the same patch.
One release, six sources, the same five flaws
The matrix puts the five side by side. The nginx.org word and the F5 scores come from brief 742's saved copies of the nginx.org advisory page and the CVE records; the rest comes from the pages saved today. Where a page covers nothing on a point, the cell is dashed. Nothing is resolved.
The rating. nginx.org's advisory page (742's copy) says major for CVE-2026-123007 and CVE-2026-113425 and medium for the other three. F5's CVE records give all five CVSS 3.1 8.1 (HIGH) and CVSS 4.0 9.2 (CRITICAL). Of the new pages, Rapid7 prints "9.2 (Critical)" under CVSSv4 for its one flaw, Wiz prints 9.2 for the first five rows of its ten row table (its text calls the 9.2 of its two flaws a CVSS 4.0 score, and the scores on its other five rows match F5's CVSS 4.0 figures in 742's copies, derived), and SecurityWeek prints 9.2 for all five. Xint prints no number: its table says Critical for all five. Censys prints no rating. The 8.1 appears only in F5's CVSS 3.1 figure, which none of the new pages quotes. A reader who meets only one of these pages meets one of these words.
How many are code execution. Xint: five. Wiz: its summary and overview say two, CVE-2026-123633 and CVE-2026-123007, and its table calls the impact of CVE-2026-123634 "RCE when attacker controls upstream host name" and of CVE-2026-113425 "RCE when configured as a mail server proxy with auth_http", while for CVE-2026-121810 it says "Exploitable configured with HTTP/3 as a proxy server". Rapid7: one, because its page covers only CVE-2026-123007. Censys: it reproduced an accounting error for CVE-2026-123007 with no crash, a use-after-free and worker crash on the stock 1.31.8 image for CVE-2026-123633, a use-after-free under AddressSanitizer plus crashes on a custom debug build but no stock crash for CVE-2026-123634, a stale-session crash confirmed as a use-after-free under AddressSanitizer for CVE-2026-113425, and did not reproduce CVE-2026-121810. It says none of the use-after-free reproductions resulted in remote code execution.
First affected release. Xint's table has a column headed "First affected release for this path". For three flaws its numbers match the start of the vulnerable range nginx.org prints (1.9.5, 1.19.7, 1.25.0); for two they do not. Wiz's table prints the same starts as nginx.org for all five. The two columns answer different questions, and the pages do not say why they differ.
First affected release as printed: Xint's column for the path it tested, and the start of the vulnerable range on nginx.org's page (742's copy), which Wiz's table repeats.
| Flaw | Xint: first affected release for this path | nginx.org and Wiz: range starts |
|---|---|---|
| CVE-2026-123007, HTTP/2 flow control | 1.9.5 | 1.9.5 |
| CVE-2026-123634, resolver with mirror | 1.13.4 | 0.6.18 |
| CVE-2026-123633, HTTP/2 with an upstream | 1.19.7 | 1.19.7 |
| CVE-2026-121810, HTTP/3 with an upstream | 1.25.0 | 1.25.0 |
| CVE-2026-113425, mail proxy with auth_http | 0.5.15 | 0.5.33 |
For the resolver flaw Xint's start is later than nginx.org's, and for the mail proxy it is earlier. Xint's own note on the resolver row says the table shows the setup it tested, which uses the mirror directive, and that other features that create subrequests, such as SSI, may reach the same bug. This site does not know what produces the other numbers. For a patch window the wider range is the safer one to plan from (this site's suggestion); the safest reading is that both are printed and neither is explained.
The conditions: ASLR, 128 GB and one CPU feature set
Rapid7's page is the one that prints conditions in numbers, and they are the ones a reader can test against their own estate. The diagram sets out what the page says its demonstration depends on, then the two statements about ASLR that no page reconciles. It is a chain of conditions, not a sequence of attack steps; the flaw's mechanism is not drawn here and the saved copy of Rapid7's page has its technical section removed.
The traffic figures the saved pages print, with this site's arithmetic (derived). 128 GB is taken as 128 times 10 to the 9th bytes unless a line says otherwise.
| Figure on a page | Arithmetic (derived) | Result |
|---|---|---|
| Rapid7: approximately 128 GB of HTTP/2 data | 128 x 10^9 bytes / 2^30 bytes per GiB | 119.2 GiB; Censys prints a range of 60 to 119 GiB for Rapid7's crash and code execution stages, so the top of its range agrees |
| Rapid7: two minutes at 10 Gbit/s | 128 GB x 8 = 1,024 Gbit; 1,024 / 10 = 102.4 s | 1.7 minutes |
| Rapid7: seventeen minutes at 1 Gbit/s | 1,024 Gbit / 1 Gbit/s = 1,024 s | 17.1 minutes (18.3 minutes if the page meant 128 GiB, so decimal fits the page better) |
| Rapid7: below about 300 Mbit/s the one hour connection lifetime is exceeded | 1,024 Gbit / 3,600 s | 284 Mbit/s (305 Mbit/s if 128 GiB); at exactly 300 Mbit/s the transfer takes 56.9 minutes, inside the hour |
| Rapid7: exploitation takes two and a half minutes in the video | 1,024 Gbit / 150 s | 6.8 Gbit/s average; the page does not print the link speed of the video |
| Censys: the whole capture is 5.69 MB | 128 x 10^9 / 5.69 x 10^6 (MB taken as 10^6 bytes) | about 22,500 times smaller (the page does not say MB or MiB) |
What this establishes and does not. On Rapid7's account, a proof of concept ran code as the nginx worker against one Docker image of 1.31.8 under the conditions above; the page does not say how many internet servers meet them. Its limits are its own: bandwidth and the connection lifetime ("practical limitations on Internet exploitation"), a proxy in the path, and the memcpy path chosen by the avx512f and avx512vl CPU features. "Alternative RCE exploitation strategies may be viable, but none are currently known", the page says, so it says nothing either way about other CPUs or other C libraries. SecurityWeek paraphrases the timing as "between two minutes and one hour"; Rapid7's page gives two minutes at 10 Gbit/s and 17 at 1 Gbit/s and treats an hour as the limit above which the attempt fails.
ASLR, as each page prints it. F5's CVE records, in 742's copies, say attackers can execute code where ASLR is disabled or the attacker can bypass it. Wiz says ASLR does not protect because attackers can use either of its two flaws for memory read access before code execution, and that its 21 October report included an ASLR bypass. SecurityWeek says Rapid7's public exploit defeats ASLR by leaking memory addresses from server responses; the saved Rapid7 text says only that the memory corruption "can disclose data", and its technical section is removed, so this site cannot check the secondary's sentence and did not fetch the proof of concept. Xint's kept text prints no ASLR statement. Censys says none of its reproductions resulted in remote code execution. The four positions are consistent with a condition that holds in some builds and not others, and consistent with a dispute; the pages do not settle which.
What the public pages and counts do not establish. Rapid7's page does not say how many servers a public proof of concept can compromise. Wiz's and Xint's claims for CVE-2026-123633, CVE-2026-123634, CVE-2026-121810 and CVE-2026-113425 are not shown on a default build on any saved page, and Censys, which tried, did not reproduce code execution. Censys's own small test (a compact cleartext HTTP/2 test, whole capture 5.69 MB, that shows different responses from vulnerable 1.31.8 and patched 1.31.9) is a patch and detection test, not an exploit; it reached the "corrupted flow-control state" without a crash.
Counting servers: a tracker with a gap its own page does not explain
The Censys tracker counts one of the five flaws, CVE-2026-123007, because it is the one that needs only HTTP/2 on; the page says the other four depend on configuration that cannot be seen from the Internet and are not counted. All figures in the table are as of the page's own "Updated 11:06 UTC" stamp on 19 November 2026, in text this site saved at 12:18 UTC. They are services (an address or hostname plus port), not servers or organisations.
Censys nginx2root tracker, as printed (updated 11:06 UTC, saved 12:18 UTC on 19 November 2026) and this site's checks (derived).
| Counter | Value | Note |
|---|---|---|
| Were exposed (since tracking began on 18 November) | 9,287,463 | Services seen on an affected version with HTTP/2 on at any point |
| Are exposed (most recent scans) | 8,823,450 | Affected version with HTTP/2 on at the most recent scan |
| Fixed since 18 November | 441,489 | Later seen on a fixed release or without HTTP/2; the same number is printed for the last 24 hours |
| Do not reveal their version | 103,456,984 | Printed as 92 percent of all detected; not shown and not judged |
| Gap between the counters (derived) | 22,524 | 9,287,463 minus 441,489 is 8,845,974, not 8,823,450; the page does not explain the difference |
| The page's 95.2 percent (derived check) | 95.2 | 8,823,450 / (8,823,450 + 441,489) = 95.2 percent; 8,823,450 / 9,287,463 = 95.0 percent |
| Version hidden share (derived check) | 91.8 | 103,456,984 / (103,456,984 + 9,287,463) = 91.8 percent, which rounds to the page's 92; its notes say these are the services that show HTTP/2 |
| Distribution builds with HTTP/2 on | 2.7 million | Printed rounded, inside the 8.8 million; Ubuntu labels its header and ships fixes without a new number, so these "may already be fixed" |
| 1.30 branch on the fix | 7.9 percent of 1,112,885 | 87,864 on 1.30.7 (derived: 7.9 percent) |
| 1.31 branch on the fix | 28.4 percent of 1,243,643 | 353,547 on 1.31.9 (derived: 28.4 percent) |
The gap is 22,524 services (derived): subtract the fixed count from the were-exposed count and the answer is 8,845,974, which is 22,524 more than the are-exposed count. The page's own 95.2 percent is computed on the smaller total, 9,264,939 (derived), so the same gap sits inside both numbers. This site's inference, labelled as one: some services that were once counted may have left the exposed group without yet counting as fixed, because the page says a service counts as mitigated only after seven days without HTTP/2. The page does not say so, and it states that figures can change as detection improves.
Different questions, different dates. Xint's page gives Shodan figures for 10 November 2026: 56.7 million hosts for product:nginx, 11.94 million (21.1 percent, derived check) with an identifiable version, and 98.9 percent of those in the range 1.9.5 to 1.31.8, about 11.8 million (derived). It says those numbers are based on reported version strings and "don't confirm that a server is vulnerable", since HTTP/2 must be on and a version may carry a backported fix. Censys requires HTTP/2 seen on the same endpoint, counts HTTP/2 seen through a load balancer or CDN, counts cleartext HTTP/2, and counts a service from a scan up to nine days old. So the 8.8 million is not a refinement of the 11.8 million: they count hosts and services, on different days, with different conditions. The 103,456,984 services that hide their version are the page's largest unknown, and it says plainly that the real number of exposed services is higher than the exposed count. The page calls itself "an estimate for situational awareness, not a source of truth".
The Linux kernel flaw: no identifier, no patch, no technique
Xint says it found a separate zero-day in the Linux kernel's SCTP subsystem and that chaining it with an nginx flaw gave root on the host, including from a Docker container running nginx. It says the bug was introduced in Linux v4.15 in 2018 and has remained for more than eight years. It says it is withholding most of the details needed to reproduce it until a kernel patch is available. The saved page prints no CVE identifier for it, no fixed kernel version and no list of affected distributions. Censys says it has not verified the kernel flaw or the chain. This site prints no technique, and the one defensive line the page prints is in the technical section below with the limits the page itself states.
What the pages state, and what they do not
The negative column is the point: for each question a reader will ask, what a saved page prints and what none of them does. Every time stamp is the page's own or the time this site saved it.
What the saved pages state and do not state, read between 12:15 and 12:19 UTC on 19 November 2026.
| Question | Stated on a saved page | Not stated on any saved page |
|---|---|---|
| Exploited in the wild? | Censys: at 2:00 p.m. Pacific Standard Time on 18 November it was not aware of any reported exploitation and none of the five was in the CISA Known Exploited Vulnerabilities Catalog; the page calls that a point-in-time absence of evidence. F5's saved index prints nothing on it; SecurityWeek says F5 makes no mention. | Anything after those moments; any report from a defender's own logs |
| How many servers can a public proof of concept compromise? | Rapid7: not through a proxy, about 128 GB of traffic, one CPU feature set, inside the connection lifetime | A number of servers; any measured success rate |
| Do the code execution claims for the other four hold on a default build? | Xint and Wiz say they lead to code execution; Censys reproduced use-after-free conditions for three, with no code execution | A default-build demonstration on any saved page; the technique sections are removed from this site's copies |
| Does any non-x86-64 or non-glibc build fall? | Rapid7: "none are currently known" | Whether any is exploitable |
| Which kernel flaw, fixed when, in which distributions? | A Linux kernel SCTP flaw introduced in v4.15 in 2018; no patch yet when Xint wrote | A CVE id, a fix date, the distributions exposed |
| How many of the 103 million version hidden services run an affected build? | They are not judged; the real exposed number is higher | Any estimate of the share |
| What identifies an attack? | Wiz: crashes or commands launched by NGINX are general warning signs, not confirmed indicators. Censys: crashes and resets do not independently indicate exploitation. Xint: little evidence may show in access logs | An indicator of compromise that identifies these flaws specifically |
| What do F5's own articles say? | The index lists ten NGINX articles and their affected products | Their text: this site did not fetch them |
Where the pages differ, left unresolved
This site reports each difference as printed and resolves none.
Which stable release. nginx.org (742's copy), Rapid7, Wiz and Censys give 1.30.7 for the stable branch and 1.31.9 for mainline. Xint's saved text names only 1.31.9 and says to update to it "when it is available"; it was written before the release.
NGINX Plus. Rapid7's remediation table lists NGINX Plus from R34 to R36 with fixed version None, and 37.0.0 to 37.0.7.1 and 37.1.0 to 37.2.0.1 fixed in 37.0.8.1 and 37.3.0.1. F5's CVE record for CVE-2026-123007 and Wiz's table both start the affected Plus range at R33. For CVE-2026-123633 Wiz prints fixed Plus releases 37.0.7.1 and 37.2.0.1, which are lower than the 37.0.8.1 and 37.3.0.1 that Rapid7 and Censys give for CVE-2026-123007; a build at the higher numbers is past both. Whether R33 is affected is a question the Rapid7 table does not answer, because it does not list it.
Workarounds. F5's records print the workaround "None" for CVE-2026-123007, CVE-2026-123633 and CVE-2026-121810. Xint, Wiz and Censys all print turning off HTTP/2 as a temporary mitigation for the HTTP/2 flaws, and Xint and Censys print turning off HTTP/3 for CVE-2026-121810. For the resolver flaw F5's record says to avoid SSI or mirror directives and variable proxied server names; Xint says setting mirror_request_body on stopped its exploit but does not fix the bug, and Censys says it does not fully resolve subrequest risks.
How the tracker describes itself. Its top line says the 103,456,984 version hidden services are 92 percent of all detected; its notes say those are the services that show HTTP/2. The arithmetic in the table matches the second reading, not a share of all nginx services.
The press copy. SecurityWeek calls Censys's test a proof of concept published on GitHub and says it warned of about 9 million affected services; Censys's own page calls it a compact test with packet captures for defenders and links a repository this site did not fetch. SecurityWeek also gives the CVSS score of 9.2 for all five and says F5 announced patches for 22 vulnerabilities, which agrees with the 22 articles on F5's index page (derived count). The figures SecurityWeek prints were checked here against a primary page wherever one exists.
What the pages print for defenders: versions, checks and one rule id
These are the defensive items the saved pages print, copied as printed and attributed in words. The saved pages print no indicators of compromise that identify these flaws specifically: Censys and Xint say crashes and resets alone are not indicators, and Wiz calls crashes and launched commands general warning signs. No trigger, frame sequence, request or exploit step is printed here: the pages print some of those, the saved copies have them removed, and this site does not reproduce them.
First the fixed versions. This is Rapid7's remediation table for CVE-2026-123007, as printed; the release notes of nginx.org name the same open source numbers.
Rapid7's remediation table for CVE-2026-123007 as printed, read on 19 November 2026.
| Product | Vulnerable versions | Fixed version |
|---|---|---|
| NGINX Plus | 37.1.0 - 37.2.0.1 | 37.3.0.1 |
| NGINX Plus | 37.0.0 - 37.0.7.1 | 37.0.8.1 |
| NGINX Plus | R34 - R36 | None |
| NGINX Open Source | 1.31.0 - 1.31.8 | 1.31.9 |
| NGINX Open Source | 1.9.5 - 1.30.6 | 1.30.7 |
Next the stop gap for the two HTTP/2 flaws. Wiz prints the instruction below; Xint prints the two lines under it for nginx 1.25.1 and later and says services that need HTTP/2, such as native gRPC, will be affected. Censys says turning HTTP/2 off covers only two of the five, and Xint says the same.
Set http2 off; in affected server blocks on NGINX 1.25.1 or later, and remove the http2 parameter from any listen directives.
listen 443 ssl;
http2 off;
Xint prints four checks for CVE-2026-123007 exposure: the version, whether the binary was built with the HTTP/2 module, whether HTTP/2 is enabled in the configuration, and whether HTTP/2 is reachable from outside. It says to run the last from a machine outside the network, and that if a CDN or reverse proxy sits in front, the test only reaches that proxy. A server is affected if it runs an unpatched version, includes the module and has HTTP/2 on, by the page's reading.
nginx -v 2>&1
nginx -V 2>&1 | grep -q -- '--with-http_v2_module' \
&& echo 'HTTP2_MODULE=present' \
|| echo 'HTTP2_MODULE=absent'
sudo nginx -T 2>&1 \
| sed 's/#.*//' \
| grep -nE 'listen[[:space:]].*http2|^[[:space:]]*http2[[:space:]]+on[[:space:]]*;'
openssl s_client -connect {HOST}:443 -servername {HOST} -alpn h2 </dev/null 2>&1 \
| grep -E 'ALPN protocol:|No ALPN negotiated|Connection refused|connect:errno|handshake failure'
Xint's four places to look for an attack: connection statistics, requests per connection, programs started by workers, and worker crashes. It says an attack "may leave little evidence in nginx access logs". Two of its checks look for results rather than for the flaw's traffic: a record of any program an nginx worker starts (its page says a normal worker has no reason to run external programs) and a search for abnormal worker exits. It warns that the audit rule records only from the moment it is installed, so an empty result cannot rule out an earlier compromise, and that a worker can crash for reasons unrelated to an attack.
ps -eo uid,user,pid,ppid,args | grep '[n]ginx: worker process'
sudo auditctl -a always,exit -F arch=b64 -S execve,execveat \
-F euid=NGINX_UID -k nginx_worker_exec
sudo ausearch -k nginx_worker_exec -ts recent -i
sudo grep -Ei 'worker process.*exited|signal 11|segfault' /var/log/nginx/error.log
sudo coredumpctl list nginx
Censys's detection note: on Suricata 8.0.1, its candidate signature ID 9007015 fired exactly once on the real vulnerable cleartext HTTP/2 capture and produced zero alerts on the patched control. The page says that validates the rule against that traffic pair, not as a general exploit detector, that its other rules are experimental, and that TLS hides the control traffic from network sensors, so it needs decrypted telemetry or cleartext connections. It prints narrow search queries; this is its host query. It searches for HTTP/2 support and an affected version in the Server header on the same endpoint, and the page says it does not check whether the required configurations are present.
host.services:(endpoints:(http.supported_versions="HTTP/2" and http.headers:(key="Server" and value=~"^nginx/((1\\.9\\.[5-9]|1\\.9\\.[1-9][0-9]+|1\\.[1-2][0-9]\\.[0-9]+|1\\.30\\.[0-6])|(1\\.31\\.[0-8]))([[:space:]].*)?$")))
For the kernel flaw Xint prints one line for hosts that do not use SCTP, to be placed in a .conf file under /etc/modprobe.d/. The page says it prevents loading only through modprobe, does not unload a module that is already loaded and does not disable SCTP compiled into the kernel. It is the page's mitigation until a kernel patch exists, not a fix.
install sctp /bin/false
The workaround words F5's own CVE records print for the five flaws (742's saved copies):
CVE-2026-123007 HTTP/2 flow control: None
CVE-2026-123633 HTTP/2 with an upstream: None
CVE-2026-121810 HTTP/3 with an upstream: None
CVE-2026-123634 Resolver with a request-derived destination: Avoid using SSI or mirror directives and avoid configuring a proxied server name with variables
CVE-2026-113425 Mail proxy with auth_http: Restrict listeners to trusted networks.
What to do, in order
The first four items are what the pages say. The rest are this site's suggestions, ordered by exposure rather than by any one rating.
Take this with you
Defender actions for the NGINX 18 November release, in order
- What the pages say: upgrade to nginx 1.30.7 or 1.31.9, or NGINX Plus 37.0.8.1 or 37.3.0.1, and restart the service; a configuration reload does not replace the running binary (Xint and Censys).
- What the pages say: distribution packages can carry the fix without a new nginx version, so compare the full package version with the distribution's advisory; Xint gives Ubuntu 22.04 and Debian 12 examples, and Censys says to confirm the advisories cover all five CVEs.
- What the pages say: disabling HTTP/2 covers only two of the five; the other three need HTTP/3 off, a fixed upstream with no request-derived destination where the resolver and subrequests are used, or a restricted mail proxy (Xint and Censys).
- What the pages say: keep nginx error logs, worker crash events and timestamps before routine rotation, and treat a program started by an nginx worker as a high priority event (Censys, Xint).
- Inventory every NGINX you run, including those inside other products: F5's index lists NGINX Instance Manager, F5 WAF for NGINX, NGINX App Protect WAF, F5 DoS for NGINX, NGINX App Protect DoS, NGINX Gateway Fabric and NGINX Ingress Controller against CVE-2026-123633 and, outside these five, CVE-2026-118899 (this site's suggestion).
- Check whether a CDN or proxy sits in front of nginx: Xint says the reachability test then reaches only the front end, and Rapid7 says a proxied connection will probably not carry the traffic its method needs (this site's suggestion on how to use both).
- On hosts that do not use SCTP, apply Xint's module line until a kernel patch exists, knowing its limits (this site's suggestion).
- Plan the patch window from your own configuration, not from one word: the five flaws are medium or major, 8.1 or 9.2, and claimed or not reproduced, depending on the page (this site's suggestion).
- Treat exploitation in the wild as unreported at the time of the pages, not as proven absent: Censys's own wording is a point-in-time absence of evidence (this site's suggestion).
The question the four accounts leave open
If the project says medium, the vendor says 9.2, one lab says root and another lab saw a crash but no shell, whose word sets your patch window, and which of your NGINX builds, HTTP/2 and HTTP/3 settings, resolver and mail proxy configurations and proxies in front would you need to check to know?
Key facts
Sources
- PrimaryXint blog post nginx2root (written before the release): five flaws, configuration table, exposure checks, response and mitigation sections, disclosure timeline, the kernel flaw's framing lines. The technical analysis sections are removed from the copy this site savedXintaccessed 2026-11-19
- PrimaryWiz Research blog post: executive summary, disclosure timeline, the ten CVE table, fixed releases and the HTTP/2 off instruction. The technical deep dive is removed from the saved copyWiz Researchaccessed 2026-11-19
- PrimaryRapid7 post on CVE-2026-123007: overview, impact and the traffic and CPU conditions, remediation table, disclosure timeline. The technical analysis is removed from the saved copyRapid7accessed 2026-11-19
- PrimaryCensys ARC post nginx2root simple test, read in full: lab results for the five flaws, detection note, actions, the host triage queryCensys ARCaccessed 2026-11-19
- PrimaryCensys nginx2root exposure tracker, living page: counters printed Updated 11:06 UTC, text saved at 12:18 UTC on 19 November 2026Censysaccessed 2026-11-19
- PrimaryF5 K000164058 Out-of-band Security Notification, 18 November 2026: an index of 22 articles; the per-CVE articles were not fetchedF5accessed 2026-11-19
- Primarynginx security advisories, as read and saved for brief 742 on 19 November 2026: the severity word and vulnerable range of each of the fivenginx.org (the NGINX project; NGINX is F5's product)accessed 2026-11-19
- PrimaryCVE-2026-123007 record (HTTP/2 flow control), as read and saved for brief 742 on 19 November 2026: CVSS 3.1 and 4.0 metrics, affected versions, workaround, ASLR wording and the CISA-ADP SSVC entryCVE Program (assigned by F5)accessed 2026-11-19
- PrimaryCVE-2026-123634 record (Resolver with a request-derived destination), as read and saved for brief 742 on 19 November 2026: CVSS 3.1 and 4.0 metrics, affected versions, workaround, ASLR wording and the CISA-ADP SSVC entryCVE Program (assigned by F5)accessed 2026-11-19
- PrimaryCVE-2026-123633 record (HTTP/2 with an upstream), as read and saved for brief 742 on 19 November 2026: CVSS 3.1 and 4.0 metrics, affected versions, workaround, ASLR wording and the CISA-ADP SSVC entryCVE Program (assigned by F5)accessed 2026-11-19
- PrimaryCVE-2026-121810 record (HTTP/3 with an upstream), as read and saved for brief 742 on 19 November 2026: CVSS 3.1 and 4.0 metrics, affected versions, workaround, ASLR wording and the CISA-ADP SSVC entryCVE Program (assigned by F5)accessed 2026-11-19
- PrimaryCVE-2026-113425 record (Mail proxy with auth_http), as read and saved for brief 742 on 19 November 2026: CVSS 3.1 and 4.0 metrics, affected versions, workaround, ASLR wording and the CISA-ADP SSVC entryCVE Program (assigned by F5)accessed 2026-11-19
- Reported bySecurityWeek, 19 November 2026: F5 patches critical NGINX2Root flaws as proof of concept exploits go public; a secondary used only to say what the press reportedSecurityWeekaccessed 2026-11-19


