NGINX: AI uncovers critical flaws hidden for years

NGINX: AI uncovers critical flaws hidden for years

An NGINX vulnerability had been in the code since 2015 and can allow unauthenticated remote code execution if HTTP/2 is enabled. On November 18, F5 released patches for ten flaws in the web server; two are critical HTTP/2 vulnerabilities discovered by several research teams with the help of artificial intelligence.

NGINX often sits at the entry point of a web architecture: it receives HTTPS requests and routes them to the relevant applications or APIs. Wiz estimates that the server powers more than 225 million websites. Because one instance can serve several services, an affected deployment can put multiple applications, websites, or APIs at risk, even if their code does not contain the flaw.

The risk lies in the layer that receives requests

The two most serious vulnerabilities, CVE-2026-123633 and CVE-2026-123007, affect the handling of HTTP/2 traffic. Both can lead to code execution on the server without requiring the attacker to authenticate. For CVE-2026-123007, Wiz demonstrated the impact on an NGINX server serving static files: no backend application or third-party modules were needed.

The specific scope depends on the configuration. CVE-2026-123007 requires HTTP/2 to be enabled and also affects NGINX servers that serve content directly. CVE-2026-123633 requires HTTP/2 and proxy traffic to an application server; the flaw occurs when NGINX closes an active connection while trying to establish a connection to that server.

Wiz describes both bugs as memory-safety issues. In CVE-2026-123633, NGINX frees the memory for a connection while a request still needs it: a use-after-free. In CVE-2026-123007, an incorrect flow-control calculation can result in an invalid memory copy. Possible outcomes include control of the NGINX process, data theft, website tampering, or service disruption.

An initiative to secure the world's software | Project Glasswing — Anthropic

What changes when AI is involved in the search

Wiz Research, Ikotas Labs, Metnir, Rapid7, and Xint independently discovered the two flaws and reported them to F5. Wiz says it used Wiz Atlas, its AI-powered code analysis system, to find them; its researchers used Gemini 4 Carbon and ChatGPT 5.6 Sol to validate the impact. The Register notes that most, if not all, of the findings by the teams mentioned involved AI assistance.

Wiz’s timeline shows how quickly some of the steps followed one another: it reported CVE-2026-123633 to F5 on October 21 and received a patch for review on November 10; it reported CVE-2026-123007 on November 16, and coordinated disclosure and the security advisory followed on the 18th. Xint said that after finding CVE-2026-123007 while preparing ZeroDay Cloud 2026, an automated scan located four more flaws without human guidance a few days later.

“AI has changed not only which vulnerabilities can be found, but how quickly longstanding risks can become real threats.”

— Tomer Schwartz, chief technology officer for exposure management and application security at Wiz

The difference is not simply that a tool suggests bugs in code. According to Wiz, the flaws required tracing subtle interactions between memory operations and different execution paths, a level of complexity that had made them difficult to discover in practice for years. AI can lower that barrier; the same change benefits defensive researchers and, potentially, those looking for vulnerabilities to exploit.

A narrower response window

Finding the flaws with AI does not eliminate the work that begins after disclosure. Organizations must identify which versions are deployed, check whether HTTP/2 is enabled, and understand whether NGINX acts as a proxy or serves content directly. In environments where one instance serves multiple applications, the inventory must cover the services sharing that entry point.

F5 released version 1.31.9 of the mainline branch and 1.30.7 of the stable branch as fixed versions for the two flaws in NGINX Open Source. NGINX Plus has different fixed versions for each CVE, so a single instruction cannot be applied to all products.

ProductFlawsFixed versions listed by Wiz
NGINX Open SourceCVE-2026-123633 and CVE-2026-1230071.30.7 stable; 1.31.9 mainline
NGINX PlusCVE-2026-12363337.0.7.1; 37.2.0.1
NGINX PlusCVE-2026-12300737.0.8.1; 37.3.0.1

If an immediate update is not feasible, Wiz and CERT-EU recommend temporarily disabling HTTP/2 on affected instances. CERT-EU proposes prioritizing systems exposed to the internet. This measure may have operational costs: organizations should check their applications’ compatibility before disabling the protocol, then re-enable it after installing the fix.

Ten advisories, with varying severity and scope

The two HTTP/2 flaws are the most prominent, but F5’s release covers ten CVEs. Other advisories include flaws related to HTTP/3, the DNS resolver, the mail proxy, and specific TLS or PROXY protocol configurations. Their exploitation conditions and impacts differ; some vulnerabilities can cause denial of service, while others require particular modules or configurations.

Severity scores differ across publications. Wiz assigns 9.2 out of 10 to the two main flaws, while CERT-EU publishes a score of 8.1 for each; the NGINX advisory classifies CVE-2026-123007 as “major” and CVE-2026-123633 as “medium.” To decide what to fix first, teams should also consider the version, configuration, and exposure of each instance, not just a single score.

CERT-EU warns that unexpected restarts of worker processes may indicate exploitation attempts, but they are not reliable proof that the server has been compromised. It also recommends reviewing process-creation logs. F5, for its part, told The Register that it had found no evidence of active exploitation at the time of its statement.

The cost of faster discovery

For security teams, the trade-off is uncomfortable: AI can help uncover old defects while also accelerating the flow of advisories that need to be assessed and fixed. Tomer Schwartz told The Register that even when a vendor releases a fix, manual work is still needed to identify vulnerable versions and configurations, test patches, deploy interim measures, and coordinate changes.

For companies running services on NGINX, the immediate action is clear: confirm which branch and version they use, check HTTP/2, and prioritize systems accessible from the internet. For software maintainers, this case shows that AI-assisted analysis can reach widely shared, long-standing components, but it does not replace coordination with the vendor or deployment to production.

Sources: theregister, Wiz, NGINX, CERT-EU