It’s NGINX’s turn to cause vulnpocalypse headaches
Security researchers disclosed five vulnerabilities Tuesday in NGINX, the web server software that handles traffic for roughly a third of all websites, each with a working exploit. The most severe allows an unauthenticated attacker to run code on the server from anywhere on the internet, and in tests the researchers chained it with a separate Linux kernel flaw to take over the machine hosting the server entirely.
F5, the company behind NGINX, released fixes Tuesday and published an advisory classifying the worst of the flaws, an integer underflow tracked as CVE-2026-123007, as critical, scoring 9.2 out of 10 on the CVSS scale.
The problem sits in HTTP/2, the protocol NGINX uses to carry many requests over a single connection. HTTP/2 uses a credit system: the client grants the server permission to send a fixed amount of data. When a client cancels a request while its response data is still waiting to be sent, NGINX hands back that unused credit without checking whether the total now exceeds the protocol’s maximum. A later check subtracts the inflated number from the maximum using unsigned arithmetic, which cannot go negative, so the calculation wraps into an enormous value. That broken number is eventually treated as a negative response length, producing a memory buffer whose end pointer sits before its start, and a memory copy that spills outside its boundaries.
Rapid7 researcher Stephen Fewer, credited in F5’s advisory along with Wiz Research, Xint.io and Ikotas Labs, built a working exploit against a NGINX 1.31.8 container with HTTP/2 and TLS enabled.
The attack has practical limits: it requires roughly 128 gigabytes of HTTP/2 traffic, most of it control frames carrying no actual data. The demonstration took two-and-a-half minutes on a fast local network, but would need about two minutes over a 10-gigabit link and 17 minutes over 1 gigabit. Because NGINX closes connections after an hour by default, a network slower than about 300 megabits per second cannot finish the attack in time. The exploit also leans on a specific optimized memory-copy routine in the GNU C Library, selected on x86-64 chips with certain AVX-512 features. Rapid7 said other exploit methods may exist but none are currently known.
The other flaws affect different NGINX features, including reverse proxying over HTTP/2 and HTTP/3, combinations of the resolver and mirror directives, and mail proxying. Wiz Research, working with Ikotas Labs, Metnir and Rapid7, independently found two of the same flaws and documented one of them, a use-after-free bug, in reverse-proxy setups where connection slots run low.
Xint, which found four of the five flaws with an automated testing harness and no human guidance, went furthest. A researcher pointed the harness at about 50,000 lines of NGINX code; 90 minutes later it produced high-severity findings, and a working exploit followed within 10 minutes. The team then chained the HTTP/2 bug with a separate flaw in the Linux kernel’s SCTP networking code to reach root, the highest privilege level on the host. F5 said its cloud services are still being evaluated while BIG-IP and F5OS products were found not vulnerable.
“When we used Xint to scan NGINX, we didn’t know what to expect, since it’s a mature, widely deployed codebase that’s undergone years of scrutiny,” Brian Pak, co-founder and CEO of Xint, said in a statement to CyberScoop. “Others have targeted NGINX with AI pentesting, but none found this critical bug, which is a testament to our team and the engineering expertise that’s gone into our AI harnesses. We’re grateful for the collaboration with project maintainers to get a patch released quickly.”
Xint’s Shodan snapshot found 98.9% of 11.94 million version-identifiable NGINX hosts running affected releases, though that counts versions, not confirmed exposure, since it does not show whether each server accepts HTTP/2 or has received a backported fix. Wiz, which validated its findings with AI models, connected the convergence of separate teams on the same long-hidden bugs to AI’s improving ability to trace memory-safety flaws, a capability available to defenders and attackers alike.
The findings come as cybersecurity researchers continue to find critical bugs in foundational software used to power internet-connected software. Just 24 hours ago, researchers disclosed a critical flaw in Java, one of the world’s most used programming languages, that could let an attacker take over servers without a password. Last month, researchers found a host of similar vulnerabilities, leading to some labeling the month as “Black October.”