Back to overview

    NGINX2Root: five flaws, ten years old, chained to root – and one affects almost every nginx

    Five nginx flaws, chained to root: the security firm Xint has disclosed, under the name NGINX2Root, a vulnerability that has been in the world’s most-used web server since HTTP/2 arrived in 2015 – ten years, actively exploited, CVSS 9.2. Chained with a kernel bug it reaches from anonymous visitor to host root. Why CVE-2026-123007 affects almost every nginx, how realistic the 128 GB exploit is, and what remains until every server is patched.

    Published by NEOSEC
    nginxNGINX2RootCVE-2026-123007HTTP/2RCEExploit-KetteKernelSchwachstellenmanagementPatch-GapXIEMNIS2Xint
    Share

    As of 19 November 2026.

    On 18 November the security firm Xint (Theori) published five vulnerabilities in nginx under the name NGINX2Root – nginx being the web server that forms the backbone of much of the internet. The headline is CVE-2026-123007: a memory corruption in HTTP/2 flow control, actively exploited, that has been in the code since nginx introduced HTTP/2 in 2015. Ten years, every official release from 1.9.5 to 1.31.8, right up to the master branch of 10 November. nginx patched on 18 November with 1.31.9 and 1.30.7.

    The name says it all: Xint did not stop at a single flaw but chained the nginx bugs with a separate Linux kernel vulnerability (SCTP) and so went from unauthenticated visitor to root on the host. For the kernel part there was no patch at publication.

    The five flaws – and why one of them affects almost everyone

    CVEComponent / bug classVulnerable whenAffected from
    CVE-2026-123007HTTP/2 flow control, memory corruptionHTTP/2 enabled1.9.5
    CVE-2026-123633Use-after-free in HTTP/2HTTP/2 enabled and forwarding to an upstream1.19.7
    CVE-2026-121810Use-after-free in HTTP/3HTTP/3 enabled and forwarding to an upstream1.25.0
    CVE-2026-123634Use-after-free in the resolverresolver set, request-derived proxy_pass, subrequests (mirror/SSI)1.13.4
    CVE-2026-113425Use-after-free in the mail proxymail proxy with proxy on and auth_http0.5.33

    The decisive one is the first. CVE-2026-123007 needs just one condition: HTTP/2 is enabled – which it is in the overwhelming majority of nginx installations. A checker service reported that 59 percent of scanned servers showed a vulnerable version. The other four require specific configurations (upstream proxying, a resolver with a variable target, the mail proxy) and hit far fewer systems in practice.

    nginx rates all of them CVSS 9.2 (CVSS 3.1: 8.1). Notable is the vendor’s own inconsistency: on the advisory page CVE-2026-123007 is called “major” and the identically scored CVE-2026-123633 “medium” – at the same score. Another reminder that a label alone does not replace prioritisation (see our assessment of Java4Shell).

    How realistic is exploitation?

    Here the sober view we owe our readers is worth it. The proof of concept Xint published for CVE-2026-123007 has, by the author’s own account, considerable prerequisites: it runs only on x86-64 servers with AVX-512 and a particular glibc memory-copy routine, has to defeat ASLR, and to do so sends roughly 128 GB of HTTP/2 traffic over a single connection. This is not a one-packet exploit.

    That bar lowers the immediate pressure – but it is no comfort. First, the flaw also suffices, without full RCE, to crash the nginx worker (denial of service). Second, this year’s history shows how quickly an unwieldy PoC becomes a reliable tool. Third, the flaw is already listed as actively exploited. The 128 GB are a property of the one published exploit, not a constant of the vulnerability.

    What to do

    1. Patch. Update to nginx 1.31.9 (mainline) or 1.30.7 (stable) and restart. That closes all five.
    2. Where immediate patching is not possible: disable HTTP/2. On 1.25.1 and later, http2 off; in each affected server block, and remove the http2 parameter from the listen lines. Important: this blocks only CVE-2026-123007 and CVE-2026-123633 – not the HTTP/3, resolver and mail-proxy paths. Those three need their own countermeasures (remove HTTP/3 listeners; point proxy_pass at a fixed upstream group; restrict the mail proxy to loopback/firewall).
    3. Rebuild containers and images. Every image that bundles nginx must be rebuilt once its maintainer ships the fix. Note: the ingress-nginx controller will not get a fix (end of life) – plan a move to another ingress controller, not a patch.
    4. Check distributions. Debian and Ubuntu ship on their own schedule; Ubuntu was still unfixed for 16.04 to 26.10 at publication, as was OpenResty. Until then the mitigation applies.
    5. Keep the kernel in view. The root chain uses a separate kernel flaw. Check host patch levels – the same pattern as in our Linux TCP post.

    Detection – because patching is a race here

    A ten-year-old flaw in the most widely used web server, found by AI, with a public PoC on disclosure day: this is the patch gap from Surveil or else in its purest form. Until every nginx in the house and at every service provider is updated, weeks pass. In that time, visibility decides.

    In XIEM, look at:

    • Crashing nginx workers (segmentation fault) – the most visible footprint of a failed or ongoing exploit attempt. Wazuh log rule on worker process … exited on signal 11 and clusters of worker restarts.
    • Anomalous HTTP/2 traffic: a single connection transferring two- to three-digit gigabytes is not explicable in any normal operation. NetFlow/reverse-proxy metrics on connection volume.
    • Web server egress: an nginx that, after a successful exploit, opens outbound connections not belonging to its upstreams is a finding.
    • Host level: new processes as the www-data/nginx user, unusual child processes of the worker master, privilege-escalation traces (the root part of the chain).

    Context

    NGINX2Root is instructive for three reasons. First, longevity: ten years in the code, through every review – evidence that “established and widely used” does not mean “audited and safe”. Second, chaining: the story is not the single flaw but the path from an anonymous HTTP request to root, wiring a web-server flaw to a kernel bug – defence must break the chain, not just close the first link. Third, the same finders: Xint/Theori already delivered a nine-year-old kernel flaw with “Copy Fail” in April; here the matching kernel piece comes from the same workshop. AI-assisted research teams are now finding such legacy bugs in series.

    For entities subject to NIS2, CVE-2026-123007 is to be treated as a priority in vulnerability management under Article 21 because of its exploitation status; if it is exploited in your environment, the 24-hour early warning under Article 23 applies.

    Sources

    Affected products