Ten patches, five grouped as nginx2root
F5, which maintains open-source NGINX, published an out-of-band security notification on Wednesday, November 18, addressing 10 CVEs. The update target is NGINX 1.31.9. NGINX handles HTTPS requests and routing for roughly a third of all websites, around 225 million sites by the counts The Register cites.
Five of those flaws are the ones Xint disclosed the same day, grouped as nginx2root. The two HTTP/2 flaws carry a verified CVSS 4.0 score of 9.2 each; AI Weekly's reporting rates all five at 9.2. Whatever the scores, individual exposure depends on configuration. A server without HTTP/2 enabled is not exposed to the two best-documented flaws. A server using the mail proxy, HTTP/3 with upstream proxying, or a resolver-plus-mirror configuration may be exposed through a different one.
What the two HTTP/2 flaws do
CVE-2026-123633 is a use-after-free: a connection's memory is freed while a request still needs it. CVE-2026-123007 is an error in HTTP/2 flow-control accounting that leads to an invalid memory copy. Wiz, which reported both independently, describes both as memory-safety issues in how NGINX handles HTTP/2 traffic.
Both allow a remote attacker to execute code without logging in, provided HTTP/2 is turned on. CVE-2026-123007 reaches back to NGINX 1.9.5, released in 2015, when NGINX introduced HTTP/2. Rapid7 has already published a proof of concept for that flaw.
Xint's other three flaws cover different configurations, per Xint's account relayed by AI Weekly: CVE-2026-123634 affects setups combining the resolver with proxy_pass and mirror directives with mirror_request_body off; CVE-2026-121810 affects HTTP/3 (QUIC) with upstream proxy connections; CVE-2026-113425 affects the NGINX mail proxy using proxy and auth_http. Xint says each of the five has a proven exploit, and that it chained one of them with a separate Linux kernel SCTP vulnerability to move from worker privileges to root on the host, including on AWS, Azure, Google Cloud and bare-metal servers. Check your own configuration against F5's advisory rather than assuming the HTTP/2 pair is the whole exposure.
The 90-minute part
The discovery method is the second story inside this one. Xint says a researcher first found CVE-2026-123007 while preparing for its ZeroDay Cloud 2026 event, and that a fully automated scan then found the other four with no human guidance. In its own account, the platform worked through a 50,000-line slice of NGINX in about 90 minutes and returned findings specific enough to become a working exploit in under ten minutes.
Reporting based on Forbes puts the cost at roughly $1,000 in OpenAI tokens, using GPT-6.1 Sol and GPT-6 Luna, models Xint notes have been generally available since May. Those are Xint's numbers, relayed by Forbes and AI Weekly, not an independent replication. Treat the 90 minutes and the $1,000 as a vendor's account of its own run.
What is independently corroborated is the convergence. Wiz says its own AI code-scanning harness surfaced the same two HTTP/2 flaws, and Rapid7, Ikotas Labs and Metnir each reported at least one of them to F5. Five teams arriving at the same small set of bugs in the same disclosure window is the part of this story that does not rest on any one firm's account.
What to do, and what disabling HTTP/2 does not do
Update to NGINX 1.31.9 on the mainline branch, or 1.30.7 on the stable branch, per reporting on F5's notice. If you cannot update immediately, disabling HTTP/2 on exposed listeners can mitigate the two HTTP/2 flaws. It does not address the other three, whose exposure depends on the resolver/mirror, HTTP/3 proxying and mail-proxy configurations described above.
Xint said it had no evidence the flaws were being exploited in the wild at publication, via Infosecurity Magazine's reporting. That sentence has a short shelf life once proof-of-concept code is public, as Rapid7's now is for CVE-2026-123007. Xint also claimed, from a November 10 scan, that 98.9% of hosts with identifiable NGINX versions were affected. Read that as a claim about hosts whose versions could be identified, not as a census of the web, and check your own build and configuration rather than assuming either way.
This lands a day after Java4Shell, a critical OpenJDK flaw also found with AI assistance. Two disclosures in two days do not measure an industry trend, but they share a shape: AI-assisted scanning produced critical flaws in foundational software, and in both cases the fix existed on disclosure day. The operational point is the plain one. The patches are out; the exposure that remains is the gap between a published fix and an applied one.
Sources
- [1] The Register, "AI found more critical bugs in another widely used technology running 225M websites," November 18, 2026 — read first-hand: F5's 10-CVE notification, the two HTTP/2 flaws and their 9.2 ratings, the 2015 reach, the five teams' independent reports, Rapid7's proof of concept.Read source
- [2] Xint via Business Wire, "PSA: Immediately Patch NGINX Web Servers Due to Critical Vulnerability Allowing Unauthenticated User RCE," November 18, 2026 — Xint's own release, read first-hand: the five-CVE nginx2root set, the 98.9% scan claim, the 90-minute autonomous discovery account, the root escalation.Read source
- [3] AI Weekly, "Xint AI Finds Five Nginx RCE Bugs; F5 Patches Wednesday," November 18, 2026 — the all-five 9.2 ratings, the other three flaws' configurations, the stable-branch target, and the Forbes-relayed token cost.Read source
- [4] Infosecurity Magazine, "Five Critical Vulnerabilities Found in NGINX," November 18, 2026 — the no-evidence-of-wild-exploitation statement.Read source