AI Found Five Critical NGINX Flaws Hiding in Plain Sight Since 2015

Human Verified

Share to: , and

A magnifying glass over glowing circuit lines with a bright crack in a dark server room lined with server racks

Security researchers have disclosed five critical flaws in NGINX, the web server that sits in front of more than 225 million websites. The flaws were found with the help of AI tools, and the worst lets an unauthenticated attacker run code on a server, and patches are already out.

The Five Flaws

All five flaws carry a critical rating of 9.2 out of 10. The worst, CVE-2026-123007, is an integer underflow in HTTP/2 flow control that corrupts memory and hands the attacker control. A second, CVE-2026-123633, is a use-after-free in connection handling. Three more affect other configurations.

The worst flaw has been hiding in the code since 2015, when NGINX first gained HTTP/2 support. Every official release from version 1.9.5 through 1.31.8 contains it. It needs only HTTP/2 enabled, and it can be exploited against a server doing nothing more exotic than serving static files.

The scale of exposure is large, though not every installation is reachable. Xint counted 56.7 million hosts running NGINX in early November, and almost 99 percent of those reporting a version fell inside the affected range. Whether a server is actually vulnerable also depends on its configuration and patches.

How AI Found Them

Wiz found the two HTTP/2 flaws with Atlas, its AI code-scanning harness, and used Gemini 4 Carbon and ChatGPT 5.6 Sol to confirm each allowed unauthenticated code execution. Xint found one flaw by hand, then a fully automated scan found the other four without human guidance.

Five teams reached the same two HTTP/2 flaws independently: Wiz, Ikotas Labs, Metnir, Rapid7 and Xint. Each reported them to F5, the company behind NGINX, which shipped fixes in coordinated disclosure. The overlap is the story — bugs that sat untouched for a decade were found five times at once.

Wiz put the shift bluntly: complexity can no longer be relied upon as a barrier to discovery or exploitation. Flaws that once needed rare expertise and months of tracing are now surfacing through automated scanning, and each new generation of models may reach further into code that humans never audited.

Our Read

This is the second AI-found critical flaw in two days. A day earlier, Calif disclosed Java4Shell, a Java flaw found with OpenAI’s Codex in about an hour. Last month, the same pattern produced PHP2Shell. What looked like isolated wins now reads as a new routine for finding old bugs.

The same capability cuts both ways. The models that help defenders find these flaws can also be pointed at them by attackers. Rapid7 has already published a working proof of concept for the worst flaw, so unpatched servers are now exposed in the open.

Why It Matters

NGINX is the front door. It accepts traffic and routes it to the applications behind it, so one compromised deployment can expose several websites and APIs at once. The flaws sit in NGINX itself, which means even applications with secure code can be reached through it.

There is no reported exploitation in the wild yet, and none of the flaws has reached the US government’s known-exploited list. That comfort is thin: each flaw has a proven exploit, the worst has a public one, and the fix is an upgrade that takes days to roll out.

Background

The pattern is now a drumbeat. Wiz’s own disclosure links this find to PHP2Shell last month and Java4Shell yesterday, and the NSA’s classified network went down this week after patching AI-found flaws. The defenders’ new tool is also the attackers’ — and the patching race is the story now.

Source: Xint’s disclosure and Wiz’s disclosure

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.