groundy
Infrastructure & Runtime

Five New nginx Memory Bugs: Patch, Disable HTTP/2, or Proxy the Origin?

Five nginx memory bugs require patching first; interim controls like disabling HTTP/2 or proxying reduce exposure but do not fix the unpatched Linux kernel SCTP flaw.

Published 8 references
On this page8 sections

On November 18, 2026, F5 and nginx.org disclosed ten nginx vulnerabilities, including five Xint-reported memory-corruption flaws collectively named nginx2root, according to Cloudflare’s disclosure account. The practical answer is to patch first, then use protocol disabling or a fronting proxy only to reduce exposure while patches roll out. Those interim controls do not cover all five flaws, and none fixes the separate Linux kernel bug used in Xint’s reported path to host root.

All five nginx2root bugs allow an unauthenticated remote attacker to corrupt memory in an nginx worker, and each carries a CVSS 4.0 score of 9.2, according to Sysdig’s analysis. The potential outcomes are a worker crash or, in Xint’s research, command execution on the server. Those are serious reasons to prioritize the fixes, but “disable HTTP/2” is not an equivalent remedy: the five flaws reach different parts of nginx.

Patch availability depends on the package channel

The releases fixing the November 18 advisories are nginx.org stable 1.30.7 and mainline 1.31.9. That version target is corroborated by Sysdig’s analysis, Cloudflare Radar’s guidance, AlmaLinux’s advisory and Tenable’s FAQ, not only by the community remediation script that automates the upgrade on Ubuntu servers.

Distribution users need to check backports rather than relying only on the upstream version string. AlmaLinux’s advisory says its Core SIG built patched packages for every nginx stream still in its RHEL life cycle, plus Kitten 10. AlmaLinux 8, 9 and 10 all shipped affected versions.

There is also a hard exception to “just patch.” AlmaLinux says retired streams, including the AlmaLinux 8 default nginx:1.14, are not receiving these fixes. On those systems, waiting for a backport is not a plan. Moving to a maintained package stream becomes part of remediation; until that happens, the interim controls below reduce only their specified attack surfaces.

Each interim control breaks a different path

Cloudflare Radar’s per-CVE guidance provides the most concrete mapping between the five flaws and temporary configuration changes. It also shows why a single HTTP/2 switch cannot stand in for patching.

FlawExposed nginx surfaceReported interim controlImportant limit
CVE-2026-123007HTTP/2 flow control; affects HTTP/2-enabled servers from nginx 1.9.5On nginx 1.25.1 and later, set http2 off; on older versions, remove http2 from listen directivesTogether with CVE-2026-123633, one change covers both HTTP/2 flaws; it does not address the HTTP/3, resolver or mail-proxy flaws
CVE-2026-123633HTTP/2 with proxy_pass to a backend and the worker low on worker_connectionsDisable HTTP/2 using the same version-dependent methodCovered by the same http2 off change; the HTTP/3, resolver and mail-proxy flaws remain
CVE-2026-121810HTTP/3 and QUIC listenersRemove QUIC listeners and set http3 offDisabling HTTP/2 alone does not address it
CVE-2026-123634Resolver path: proxy_pass with variables plus an SSI or mirror subrequest resolving the same namePoint proxy_pass at fixed upstream groupsSetting mirror_request_body on is explicitly not a fix
CVE-2026-113425nginx mail proxyLimit mail proxy listeners to trusted networksHTTP/2 and HTTP/3 listener changes do not address it

After changing these listener settings, Radar instructs operators to restart nginx so existing connections close. A configuration that protects new connections is not enough if already-established sessions remain open.

The validation status matters. Sophos’s advisory relays the reported HTTP/2 and HTTP/3 mitigations but says Sophos has not validated them. It advises testing before deployment and treating them as temporary. It also confirms that they leave the other vulnerabilities unaddressed. These controls are therefore a rollout-window measure, not an alternative patch strategy.

CVE-2026-123007 deserves particular attention during triage. Radar describes it as an HTTP/2 flow-control overflow causing out-of-bounds heap reads and writes, with applicability reaching back to nginx 1.9.5 whenever HTTP/2 is enabled. Xint demonstrated command execution against a public HTTPS endpoint on AWS, GCP and Azure, Radar reports. That is researcher-reported evidence of feasibility, not evidence that every HTTP/2 deployment is exploitable in the same way.

When does proxying the origin actually help?

A fronting proxy can help because it changes which process parses the attacker’s protocol traffic. Xint’s nginx2root site says that for the HTTP/2 vulnerabilities, a frontend can block the attack path when it terminates HTTP/2, forwards requests to nginx over HTTP/1.1, and blocks direct access to the nginx server. The origin then receives requests through a different protocol path instead of parsing the exposed HTTP/2 connection itself. Xint also names the arrangement’s two failure modes: it does not help if the proxy does TLS passthrough, or if the proxy-to-origin connection still speaks HTTP/2.

The third condition is essential. Terminating HTTP/2 at a proxy does little for the demonstrated origin path if clients or attackers can bypass that proxy and connect directly to nginx. The architectural question shifts from “is the origin patched?” to “can anything reach the origin directly and speak the affected protocol to it?” Public direct-to-origin routes therefore deserve the same triage as the listener configuration.

This advice needs two boundaries. First, Cloudflare also advocates proxying in its vendor material, which is self-interested because Cloudflare sells proxy services; Xint’s independent endorsement of the HTTP/2-terminating arrangement makes the recommendation stronger, but does not broaden its scope. Second, Xint’s recommendation specifically addresses HTTP/2. The evidence supports disabling HTTP/3 for CVE-2026-121810, but it does not establish that every proxy arrangement automatically covers HTTP/3, the resolver path or the mail proxy.

Patching nginx does not patch the reported kernel chain

The “root” in nginx2root involves more than the nginx worker. Sysdig reports that Xint chained one nginx bug with a previously unknown Linux kernel SCTP flaw to gain root on the host, including from inside a Docker container running the official nginx image.

The kernel issue is described as a use-after-free and double-free in SCTP, introduced by Linux 4.15 in 2018. According to Xint’s account as relayed by Sysdig, it can leak kernel memory to defeat kernel address space layout randomization and allow an attacker to call an arbitrary kernel function. As of the disclosure-week writeup, it had no CVE and no patch, and researchers were withholding further details until a fix existed.

That distinction changes what remediation promises. An nginx patch addresses the nginx flaws; it does not repair the underlying SCTP kernel defect. Nor do HTTP listener changes or an HTTP/2 frontend constitute a kernel fix. The evidence does not establish another working host-root chain after the nginx entry point is patched, so it would be equally wrong to claim that the unpatched kernel bug leaves every patched nginx host remotely exploitable.

Radar offers a narrow stopgap for hosts that do not use SCTP: adding install sctp /bin/false under /etc/modprobe.d to prevent the module from loading. It does not unload a module that is already loaded and does not disable SCTP support built directly into the kernel. It is therefore a host-specific reduction step with explicit limits, not a general remediation for the container-escape chain.

Nine million services is not nine million confirmed targets

Censys’s exposure estimate, reported by Tenable, put roughly nine million nginx services on affected versions with HTTP/2 enabled as of November 18. Tenable cautions that the count includes distribution builds that may already carry the fix.

Use that figure as a reason to inventory public listeners and package provenance, not as a count of unpatched or exploitable machines. The useful fleet questions are more specific: which hosts still accept HTTP/2 or HTTP/3 directly, which package stream supplies nginx, whether that stream has received a backport, and whether a resolver or mail-proxy path remains exposed.

Circulating exploit claims can be judged against documented constraints. Tenable’s FAQ reports that Rapid7’s public proof of concept for CVE-2026-123007, published at sfewer-r7/CVE-2026-123007 on GitHub, was demonstrated against nginx 1.31.8 in a Docker container with HTTP/2 over TLS. Tenable relays Rapid7’s description of it as constrained: roughly 128 gigabytes of HTTP/2 traffic on one connection, a dependence on a glibc memory-copy routine specific to x86-64 builds, ASLR defeated by recovering heap addresses from server responses, execution landing only in the nginx worker, and unlikely success behind HTTP/2-terminating proxies or on connections too slow to deliver the data before nginx’s one-hour connection limit. Those constraints form a checklist for judging a claimed exploit: expect a very large single-connection transfer, worker-level execution rather than host root, and failure behind an HTTP/2-terminating proxy. A claim of immediate host root is describing the SCTP chain, which still depends on the unpatched kernel flaw. Tenable also states that the five CVEs were not on CISA’s Known Exploited Vulnerabilities catalog as of November 19 and that no source it reviewed reports exploitation in the wild; Cloudflare’s November 20 post reports the same dated status and notes that fake “exploits” are circulating on GitHub. The upgrade script linked above is a remediation tool, not an exploit.

I would patch supported nginx installations first, using the fixed upstream releases or an appropriate distro backport. During the rollout, I would disable the affected public HTTP/2 and HTTP/3 listeners where the service can tolerate it, or use the HTTP/2-terminating frontend arrangement with direct origin access blocked. I would handle the proxy_pass and mail-proxy guidance separately, restart nginx after the listener changes, and avoid representing any of those steps as coverage for all ten advisories or as a fix for the unpatched SCTP kernel flaw.

Frequently Asked Questions

Which nginx versions contain the fixes for the nginx2root vulnerabilities?

The releases fixing the November 18 advisories are nginx.org stable 1.30.7 and mainline 1.31.9.

Does patching nginx fix the Linux kernel SCTP flaw?

An nginx patch addresses the nginx flaws; it does not repair the underlying SCTP kernel defect.

References

Follow the links in the article for context. The supporting material is collected here for further reading.

  1. Cloudflare’s disclosure accountblog.cloudflare.comAccessed
  2. Sysdig’s analysissysdig.comAccessed
  3. Cloudflare Radar’s guidanceradar.cloudflare.comAccessed
  4. AlmaLinux’s advisoryalmalinux.orgAccessed
  5. Tenable’s FAQtenable.comAccessed
  6. community remediation scriptgithub.comAccessed
  7. Sophos’s advisorysophos.comAccessed
  8. nginx2root sitenginx2root.comAccessed

Join the discussion

Share a useful perspective or ask a question about this article.

Discussion guidelinesComments privacy