AI-Fueled Nginx HTTP/2 Zero-Days Hit 225M Sites: 11-Year Flaws, NGINX2Root Chain
Five research teams independently discovered two critical HTTP/2 vulnerabilities in Nginx (CVE-2026-123633 and CVE-2026-123007) using AI tools, affecting versions from 1.9.5 to 1.31.8 and enabling unauthenticated RCE on 225 million sites; a full exploit chain dubbed NGINX2Root combines five Nginx RCEs with an unpatched Linux kernel SCTP flaw to achieve container escape and host root access on major clouds.
Overview
On November 18, F5 released security updates addressing 10 CVEs in Nginx. Two critical vulnerabilities reside in the HTTP/2 implementation: CVE-2026-123633 (use-after-free) and CVE-2026-123007 (integer underflow), both scoring CVSS 4.0 9.2 and allowing unauthenticated remote code execution. Nginx powers over 225 million websites, typically sitting as a reverse proxy in front of all applications and APIs; a compromise at this layer bypasses perimeter defenses regardless of backend code security.
Technical Details of the Vulnerabilities
CVE-2026-123007: HTTP/2 Flow Control Integer Underflow
HTTP/2 uses a flow-control window to limit data the sender may transmit before receiving acknowledgments. Nginx tracks this quota with a send_window counter. When a stream is cancelled, the unsent data quota is returned to the counter, but the code omitted the protocol-mandated upper-bound check. An attacker floods WINDOW_UPDATE frames to inflate the counter; subsequent bounds checks fail due to unsigned integer underflow, eventually triggering an out-of-bounds memory copy in the TLS send path. This can leak memory, corrupt the heap, and achieve RCE.
The flaw spans from Nginx 1.9.5 (2015, when HTTP/2 was introduced) through 1.31.8 — over a decade in the codebase. Wiz Research demonstrated that even a default configuration serving only static files with HTTPS and HTTP/2 enabled is exploitable, and ASLR does not prevent exploitation.
CVE-2026-123633: HTTP/2 Connection Use-After-Free
Nginx maintains a reusable connection list. When an HTTP/2 connection receives only the first half of a HEADERS frame, it is marked reusable; when the remaining frames arrive and request processing begins, the reusable flag is not cleared. Under connection-slot pressure, Nginx may close and free a connection that is still actively processing a request. Subsequent request-handling code continues operating on the freed memory, creating a classic use-after-free condition. This vulnerability exists since 1.19.7 (2021) and primarily affects reverse-proxy deployments that forward requests to upstream servers.
Both vulnerabilities require HTTP/2 to be enabled; servers running only HTTP/1.1 are not affected.
Discovery and Disclosure Timeline
At least five independent research teams — Wiz Research, Ikotas Labs, Metnir, Rapid7, and Xint — discovered the same vulnerabilities. Wiz reported CVE-2026-123633 on October 21 with a full ASLR-bypassing RCE proof-of-concept and fix suggestions; F5 acknowledged another team had already reported the same issue. Key dates:
Oct 21: Wiz submits CVE-2026-123633 report; F5 confirms prior report.
Nov 10: F5 provides patch for researcher review.
Nov 16: Wiz reports second vulnerability ( CVE-2026-123007).
Nov 17: F5 confirms inclusion in next-day security release.
Nov 18: Coordinated public disclosure and fixed versions published.
Critically, the fix code was merged into the repository a week before public disclosure. For attackers, the diff serves as a vulnerability specification, shrinking the effective response window. Downstream distributions (Ubuntu, Debian, RHEL) often lag in packaging, leaving many "fully updated" systems exposed.
NGINX2Root Attack Chain
Xint Security Research (under Theori) disclosed an exploit chain named NGINX2Root that escalates from Nginx worker compromise to host root and Docker escape. The chain combines five Nginx RCE vulnerabilities — the two HTTP/2 flaws plus DNS resolver UAF ( CVE-2026-123634), HTTP/3 UAF ( CVE-2026-121810), and mail proxy UAF ( CVE-2026-113425) — with a Linux kernel SCTP subsystem privilege escalation vulnerability present since kernel 4.15 (still unpatched). Testing confirms the chain works on AWS, GCP, and Azure.
Temporary mitigation for the kernel component: disable SCTP module loading via install sctp /bin/false on hosts that do not require SCTP.
Mitigation and Remediation
Version Upgrades
Fixed releases are branch-specific: 1.30.7 (stable) and 1.31.9 (mainline). Note that versions 1.31.0 through 1.31.8 are not fixed despite higher version numbers. Branches older than 1.30 have no official patches and must be upgraded.
Container images and Ingress Controllers bundle their own Nginx copies. NGINX Ingress Controller is fixed in 5.8.0; its 2026-lts-r11 release still ships vulnerable 1.31.0.
Temporary HTTP/2 Disable
If immediate upgrade is impossible, disable HTTP/2: remove the http2 parameter from listen directives; for Nginx 1.25.1 and later, also add http2 off; inside the server block. This only mitigates the two HTTP/2 CVEs; the other eight CVEs in the release are unrelated to HTTP/2.
Operational Pitfalls
Reload is insufficient. A graceful reload leaves old worker processes serving established HTTP/2 long-lived connections (shown as worker process is shutting down in ps output). Only a full restart terminates those connections.
CDN/WAF is not a silver bullet. Protection only materializes if the CDN terminates HTTP/2 and proxies to the origin via HTTP/1.1. If the CDN passes HTTP/2 frames through, mitigation depends on the specific implementation and should not be relied upon.
Detection and Monitoring
Administrators should watch for abnormal Nginx worker crashes and unexpected commands spawned by Nginx processes. Exploitation of CVE-2026-123007 requires maintaining long-lived HTTP/2 connections while sending large volumes of WINDOW_UPDATE frames; detection rules targeting this pattern can provide early warning. As of disclosure, Rapid7 has published a PoC for CVE-2026-123007 on GitHub; CERT-EU reports no observed wild exploitation, but the combination of public PoC, 225 million exposed sites, and distribution patch lag makes scanning and attempted exploitation imminent.
Conclusion
This event — decade-long dormancy, five independent discoveries, a patch merged a week before disclosure, and a full worker-to-root escape chain — is not an isolated incident. It follows the pattern of "foundational component chain explosions" seen earlier this year with PHP2Shell and Java4Shell. As AI drives the cost of discovering memory-safety vulnerabilities toward zero, the assumption that "our code has always been stable" ceases to be a valid security posture. For security teams, the priority shifts from merely chasing patches to shortening asset identification and attack-surface reduction cycles, because the next diff may already be sitting in a repository.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
