Nginx Fixes Vulnerabilities Discovered with AI

F5 has released Nginx versions 1.31.9 and 1.30.7 to address two HTTP/2 vulnerabilities along with eight other flaws, several of which were discovered using AI agents. The most severe vulnerability allows for remote code execution without authentication. The U.S. agency CISA has not yet reported any exploitation of these two vulnerabilities in the wild.
Verify that the patches are applied
A version number alone is not enough to establish that a server is protected, warns Daniel Shechter, co-founder and CEO of Miggo Security: some distributions incorporate fixes via backporting without changing the version number. A simple reload does not replace the executed code either, he emphasizes. In his view, the main risk lies with teams that might mistakenly believe they are protected, more so than the 11 years during which the bug went unnoticed. He recommends treating the server as exposed until the effective execution of the patch is proven.
For Nginx Plus, the recommended versions are 37.3.0.1 on the new branch and 37.0.8.1 for the 37.0 branch. The R33 to R36 branches, now out of support, will remain unpatched. F5 advises immediate updating or disabling HTTP/2 on the affected ports. The latter measure blocks the two most severe vulnerabilities but notably interrupts native gRPC.
It does not neutralize CVE-2026-123634, CVE-2026-121810, and CVE-2026-113425, the other discoveries from Xint, which require specific configurations. The resolver vulnerability affects installations that choose their destination server with the mirror directive for each request; the HTTP/3 vulnerability concerns QUIC listeners; and the mail proxy vulnerability involves auth_http. Xint has released mitigation measures for each.
Wiz recommends examining abnormal terminations of Nginx worker processes as well as the commands they execute. Daniel Shechter also warns that the HTTP/2 traffic exploited by the attack evades both WAFs and access logs.
Functional exploits under certain conditions
Shodan counts 56.7 million Nginx hosts; among those that expose their version, 98.9% belong to the vulnerable range. All research teams have produced a functional exploit, and two consider the exploitation easy. Xint claims that an externally reachable TLS/HTTP/2 port is sufficient to corrupt the send window, reveal addresses, and then execute commands. The company states it has validated its attack on public IPs at AWS, GCP, and Azure, in the presence of concurrent traffic.
At Metnir, exploiting CVE-2026-123633 took about 10 minutes with ASLR enabled, most of the time spent bypassing this memory address randomization. The team claims to have reproduced the use of freed memory in OpenResty and caused crashes in the distributed binaries of Kong, Apache APISIX, NGINX Gateway Fabric, and community ingress-nginx. These software packages enable HTTP/2 by default.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Stephen Fewer, a senior security researcher at Rapid7, estimates the volume required for his exploit at around 128 GB of HTTP/2 data. He estimates their transmission at about two minutes with 10 Gbit/s, compared to 17 minutes with 1 Gbit/s. His technique also relies on a glibc copy routine reserved for processors with AVX-512, common extensions in servers.
Yordan Ganchev, a senior threat intelligence specialist at watchTowr, tempers the announced ease: according to him, the target must be directly accessible, without a WAF or CDN ensuring HTTP/2 termination. It is also necessary to organize the memory heap and adapt the exploit to the versions, builds, and configurations of Nginx. On an exposed server under load, he warns, attempts may lead to more crashes than successful code executions.
Multiple teams discover the same flaws
Nginx, which is estimated to hold a third of the Internet, presents with CVE-2026-123007 a flaw that could threaten millions of sites. The mechanism relies on manipulating the windows regulating HTTP/2 flows, potentially causing an out-of-bounds memory copy. Xint.io claims to have achieved command execution with the rights of a Nginx worker process through this means. All official versions since 1.9.5, which introduced HTTP/2 in 2015, are affected, spanning over ten years.
Rapid7, Wiz, Xint, and Ikotas Labs reported this vulnerability to F5. The other HTTP/2 flaw, CVE-2026-123633, called Drainbreak, corresponds to a use-after-free memory issue in versions 1.19.7 to 1.31.8. Metnir, Wiz, and Xint discovered and reported it.
These discoveries overlap despite independent research utilizing AI. Stephen Fewer does not mention this usage in his report; the other four teams employed agents or AI platforms. Xint claims five discoveries thanks to its autonomous AI, and its CEO Brian Pak states that the platform scanned 50,000 lines in 90 minutes to identify a critical vulnerability allowing remote execution without authentication. Other flaws have recently been spotted with AI in PHP, Node, OpenSSH, and Java.
An associated Linux flaw remains unpatched
Xint also claims to have discovered a vulnerability in SCTP, a subsystem of the Linux kernel, which can be combined with an Nginx flaw to achieve root privileges. The company has dubbed this combination nginx2root. During its testing, it claims to have breached the isolation of a Docker container running Nginx and then obtained root rights on the host machine.
The vulnerability has existed since Linux 4.15, released in 2018, and still lacks a CVE or a patch. The company is keeping its details confidential while waiting for distributions to provide a fix. In the meantime, administrators not using SCTP can prevent its loading by adding "install sctp /bin/false" to a file in the /etc/modprobe.d directory.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.