Nginx, which is estimated to power a third of the internet, has shipped fixes for a critical flaw that lets an unauthenticated attacker run code on the server. The bug has been in the codebase for over a decade. The patches matter, but the more interesting story is how the bugs surfaced. Several security teams independently found some of the same flaws using AI agents, which adds to a growing list of AI-discovered vulnerabilities in PHP, Node, OpenSSH, and Java.
The HTTP/2 bugs at the top of the list
The worst of them, CVE-2026-123007, abuses flow-control windows in Nginx's HTTP/2 implementation to trigger an out-of-bounds memory copy. Researchers at Xint.io turned that into command execution with Nginx worker privileges. Every official release going back to 1.9.5, the version that added HTTP/2 support in 2015, is affected.
A second HTTP/2 flaw, CVE-2026-123633, is a use-after-free affecting releases 1.19.7 through 1.31.8. Both bugs are fixed in Nginx 1.31.9 and 1.30.7, which F5 released alongside fixes for eight other vulnerabilities.
Xint's team used its own autonomous AI harness to find five of the flaws. It also turned up a separate Linux kernel bug in the SCTP subsystem. Chained with any of the Nginx flaws, that kernel bug can yield root on the server. Xint has dubbed the whole collection nginx2root. In its tests, the chain escaped a Docker container running Nginx and reached root on the underlying host.
The kernel flaw dates to Linux 4.15, released in 2018. It has no CVE and no patch yet, so Xint is holding back details until distributions ship a fix. That's the part I'd watch closely, because a container boundary is exactly what many teams assume is protecting them.
Who found what, and how
The critical HTTP/2 flaw was reported to F5 by Rapid7, Wiz, Xint, and Ikotas Labs. The second, dubbed Drainbreak, came from Metnir, Wiz, and Xint.
Only Rapid7's Stephen Fewer, a senior principal security researcher, doesn't mention AI in his write-up. The other four teams relied on AI agents and harnesses. Xint CEO Brian Pak said his platform scanned "50,000 lines of code in just 90 minutes" and surfaced "a critical vulnerability that gave an unauthenticated user a RCE path — essentially an internet breaker."
That speed is the real signal here. Ninety minutes against a codebase that has been mature for more than a decade suggests that human review has been losing ground on this class of bug for a while.
Exploitation is likely, even without confirmed attacks
The US Cybersecurity and Infrastructure Security Agency has no evidence that either flaw has been exploited in the wild. Still, every research team built a working exploit, and two described the flaws as readily exploitable.
Xint rates CVE-2026-123007 as highly exploitable. An externally reachable TLS/HTTP/2 port was enough to corrupt the send window, disclose memory addresses, and execute commands. The team validated the attack against public IPs on AWS, GCP, and Azure under concurrent traffic.
Metnir's exploit for CVE-2026-123633 worked with ASLR enabled and took about 10 minutes, nearly all of it spent defeating memory layout randomization. It reproduced the use-after-free in OpenResty and crashed the shipped binaries of Kong, Apache APISIX, NGINX Gateway Fabric, and community ingress-nginx. All of those enable HTTP/2 by default.
Fewer estimated that his exploit requires sending roughly 128GB of HTTP/2 data to the target. That takes about two minutes on a 10Gbit/s network and 17 minutes on 1Gbit/s. The technique also depends on a glibc memory-copy routine that runs only on CPUs with AVX-512 extensions, though those are common on servers.
Not everyone thinks weaponization will be simple. "Exploitation is by no means a walk in the park, as the vulnerable systems must be directly exposed to the internet with no WAF or CDN terminating HTTP/2," said Yordan Ganchev, principal threat intelligence specialist at watchTowr. "Memory corruption can also be tricky, as an effective exploit must properly groom and shape the memory heap and handle different NGINX versions, builds, and configuration variants. And when the system is internet-facing and under load, exploitation can cause more crashes than successful code execution."
Ganchev's caveats are fair, but they're a reason for caution, not comfort. Crashes on internet-facing servers are already an outage, and a determined attacker only needs one clean hit.
Patch first, then verify what's running
Nginx Plus customers should move to 37.3.0.1 on the newer release line or 37.0.8.1 on the 37.0 line. The older R33 through R36 line won't receive a fix because those versions have reached the end of technical support. If you're still on them, the upgrade is the only real option.
F5 recommends updating immediately or disabling HTTP/2 on affected ports. Disabling it blocks the two most severe flaws, but it has a cost: services that depend on HTTP/2, such as native gRPC, stop working. The other three Xint findings sit in different components and need their own handling. The resolver flaw affects setups that choose backends per request and use the mirror directive. The HTTP/3 flaw hits QUIC listeners. The mail proxy flaw affects deployments that use proxy on with auth_http. Xint published per-flaw mitigations in its technical write-up.
For the kernel flaw, there's a stopgap. If you don't use SCTP, you can block the module from loading by adding install sctp /bin/false to a file under /etc/modprobe.d. Wiz also advises checking for unexpected Nginx worker crashes and for commands launched by Nginx workers, since those are the fingerprints a successful exploit would leave.
Daniel Shechter, co-founder and CEO of Miggo Security, sees a bigger problem than the bugs themselves. "The real risk with this nginx bug isn't the 11 years it went unnoticed, it's how many teams will think they've patched when they haven't," he said. "The version number can be wrong because distros backport fixes without changing it, a reload doesn't swap out the running code, and the attack happens in HTTP/2 traffic that WAFs and access logs never see. If you can't prove the fix is actually running, assume you're still exposed."
That's my read too. Check the binary that's actually serving traffic, confirm the worker processes have restarted onto it, and test whether HTTP/2 is reachable from the internet. A version string in a config file won't tell you any of that.