NGINX fixes HTTP/2 flaws, KVM fixes nested SVM: what to check for CVE-2026-123007 and CVE-2026-99312
It has been one vulnerability after another for the past few weeks: Apache’s HTTP/2 on 11 November, PostgreSQL’s 52 on the 12th, OpenJDK’s Base64 on the 17th, and two more on the 18th. NGINX released 1.31.9 and 1.30.7 to fix CVE-2026-123007 and CVE-2026-123633 in HTTP/2. Neither needs a login, and F5 scores both 9.2 under CVSS 4.0. On the same day the Linux kernel published CVE-2026-99312, where KVM’s nested virtualisation on AMD hosts can leave an L1 guest running on a stale MMU state.
Update both. Move NGINX to 1.31.9 or 1.30.7 and restart it, and if your packages are not ready, turn HTTP/2 off for now. Move AMD hosts running KVM to 6.12.118, 6.18.64 or 7.3.5, and switch nested off wherever nobody uses it. As of 08:25 Taipei time on 19 November, Debian and Ubuntu had not fixed their nginx packages, CISA KEV listed none of the three CVEs, and I had found no report of exploitation in the wild. Rapid7 and Wiz both say they have built a PoC that executes code, and Rapid7’s write-up links to a PoC repository.
Correction (19 November 2026, 10:22 Taipei): added CVE-2026-118899, which does not need HTTP/2, and noted that Rapid7’s PoC has a public link. Suggested by Sheng, checked by Claude (claude-mythos-5-2).
Who is in scope
| NGINX CVE-2026-123007 | NGINX CVE-2026-123633 | KVM CVE-2026-99312 | |
|---|---|---|---|
| Affected (upstream) | 1.9.5 to 1.31.8 | 1.19.7 to 1.31.8 | kernel 6.1 onwards, plus stable 5.15.81 and 6.0.11 onwards |
| Conditions | HTTP/2 enabled | HTTP/2 enabled, requests proxied to a backend, worker_connections nearly used up | AMD host with KVM nested virtualisation (nested SVM) on |
| Fixed in | 1.31.9, 1.30.7 | 1.31.9, 1.30.7 | 6.12.118, 6.18.64, 7.3.5 |
| Score | CVSS 3.1 8.1, 4.0 9.2 | CVSS 3.1 8.1, 4.0 9.2 | none in the CVE record; Red Hat gives 5.6 |
NGINX Plus needs 37.0.8.1 or 37.3.0.1 to carry both fixes; CVE-2026-123633 alone is fixed in 37.0.7.1 and 37.2.0.1. nginx.org itself rates 123007 as major and 123633 as medium, and F5 says code execution needs ASLR to be disabled or bypassed. Wiz says it did bypass ASLR on the official 1.31.7 and 1.31.8 builds, on a server with no backend that only served static files. Intel hosts are not affected by CVE-2026-99312, because, as the CVE record puts it, INIT is always intercepted on VMX. The NGINX versions in the table are upstream ranges, so check your distribution’s own notice for backports.
This NGINX release fixes ten security issues in all, and turning HTTP/2 off only deals with two of them. The one with the widest reach is CVE-2026-118899, which nginx.org rates major and F5 scores 7.5 under CVSS 3.1. All it needs is listen with ssl and TLS 1.3 enabled: an attacker can send a worker into an infinite loop while it reads the request body, so it stops handling connections, a denial of service. In my test, listen 443 ssl on nginx:1.31.9 with no ssl_protocols line negotiated TLS 1.3, so most ordinary HTTPS sites meet the condition; I did not reproduce the flaw itself. There is also a use-after-free in HTTP/3 (CVE-2026-121810) and one each in the resolver and the mail proxy. Turning HTTP/2 off is a stopgap, and the update still has to happen.
What went wrong
CVE-2026-123007 is in HTTP/2 flow control. Each connection keeps a send window recording how much data it may still send. CHANGES describes that window overflowing, and F5 says the result is a heap buffer overflow that usually ends in a worker segmentation fault. Rapid7 adds some practical limits: its attack has to send a great deal of data, about two minutes on a 10 Gbit/s link and seventeen on 1 Gbit/s, and anything slower than roughly 300 Mbit/s runs past NGINX’s default one-hour connection lifetime. Rapid7 says its PoC generally does not work through another HTTP/2 proxy, and that any code it runs has the worker’s privileges. Those limits describe Rapid7’s own PoC and may not hold for other approaches.
CVE-2026-123633 is a use-after-free. When NGINX runs short of connection slots it reclaims idle connections, and Wiz explains that an HTTP/2 connection still handling a request can be treated as reclaimable, so the request goes on using memory that has been freed. That is why worker_connections running low appears in the conditions.
CVE-2026-99312 arises when KVM emulates an x86 INIT. It used to reset the MMU only if the guest had paging on, but the MMU state also records whether L2 (a guest inside a guest) is running. If L2 receives an INIT with paging off, KVM goes back to L1 still holding the stale MMU state. The fix is short: reset on every INIT. The CVE record does not say what the concrete impact is and gives no score. Red Hat rates it Moderate at 5.6 and describes the outcome as denial of service.
Checking the mitigation on OrbStack
The test below uses ordinary curl requests to see what “turn HTTP/2 off and reload” really does. It sends no attack traffic and does not reproduce the vulnerability. To avoid certificates it uses cleartext h2c:
# Ordinary curl requests only: what happens after HTTP/2 is turned off. No attack traffic.
docker network create h2n >/dev/null
docker run --detach --name ng --network h2n nginx:1.31.9 sh -c '
head -c 600000 /dev/zero > /usr/share/nginx/html/big.bin
printf "events {}\nhttp { server { listen 80; http2 on; root /usr/share/nginx/html;\n location /slow { limit_rate 2k; alias /usr/share/nginx/html/big.bin; } } }\n" > /etc/nginx/nginx.conf
exec nginx -g "daemon off;"' >/dev/null
sleep 2
c() { docker run --rm --network h2n curlimages/curl:8.17.0 --silent --output /dev/null --max-time 5 --write-out "$1 -> HTTP/%{http_version} %{http_code}, curl exit %{exitcode}\n" "${@:2}"; }
workers() { docker exec ng sh -c 'for p in /proc/[0-9]*; do tr "\0" " " < $p/cmdline; echo; done' | grep '^nginx'; }
c "1. http2 on" --http2-prior-knowledge http://ng/
# Leave one very slow HTTP/2 download open, then turn HTTP/2 off and reload
docker run --detach --name slow --network h2n curlimages/curl:8.17.0 --silent --output /dev/null --http2-prior-knowledge http://ng/slow >/dev/null
sleep 2
docker exec ng sh -c 'sed -i "s/http2 on;/http2 off;/" /etc/nginx/nginx.conf && nginx -t 2>&1 | tail -1 && nginx -s reload 2>/dev/null'
sleep 2
c "2. http2 off, reload" --http2-prior-knowledge http://ng/
workers
docker rm --force slow >/dev/null; sleep 3
echo "-- after that old connection ends"; workers
# The old-style http2 parameter is still on listen
docker exec ng sh -c 'sed -i "s/listen 80;/listen 80 http2;/" /etc/nginx/nginx.conf && nginx -s reload 2>/dev/null'
sleep 2
c "3. listen 80 http2 + http2 off" --http2-prior-knowledge http://ng/
docker rm --force ng >/dev/null; docker network rm h2n >/dev/null
Actual output (Docker 29.8.2, 19 November 2026):
1. http2 on -> HTTP/2 200, curl exit 0
nginx: configuration file /etc/nginx/nginx.conf test is successful
2. http2 off, reload -> HTTP/0 000, curl exit 16
nginx: master process nginx -g daemon off;
nginx: worker process
nginx: worker process is shutting down
-- after that old connection ends
nginx: master process nginx -g daemon off;
nginx: worker process
3. listen 80 http2 + http2 off -> HTTP/2 200, curl exit 0
Two things stand out. After the reload, new connections can no longer negotiate HTTP/2, but the old worker that still holds a connection stays around, marked is shutting down, and that old HTTP/2 connection keeps running until it finishes. Wiz advises checking that the old workers have exited, and this is why; it matters most on services with long-lived connections.
The second is easier to trip over. If the old-style http2 parameter is still on listen, the listener went on answering in HTTP/2 even with http2 off; in the server block, and nginx -t only printed a deprecation warning. The http2 directive arrived in 1.25.1, and config files carried forward from older versions often have both forms, so change both when you turn it off. I repeated the test over TLS with a self-signed certificate and listen 443 ssl http2; and got the same result: http2 off; on its own gave HTTP/1.1, and with the parameter left on listen it was HTTP/2 again.
The same run also looked at the official image defaults. nginx:latest is already 1.31.9 and nginx:stable is 1.30.7; the default config does not enable HTTP/2, and worker_connections is 1024. That is only the starting point, so check the deployed config for http2 even when you use the official image. After updating the package, restart: a reload re-reads the config but keeps running the old binary.
The KVM side cannot be checked on OrbStack. Its VM is aarch64 on kernel 7.2.12-orbstack, with no /dev/kvm and no kvm_amd module. The commands below are for an AMD host running KVM, and I have not run them:
# On an AMD host: is nested virtualisation on? Y or 1 means it is
cat /sys/module/kvm_amd/parameters/nested
uname -r
The kernel documentation says nested has been on by default since 4.20, though a distribution may change that. To turn it off, add options kvm-amd nested=0 to a .conf file under /etc/modprobe.d/, stop every VM that uses KVM, unload and reload the kvm_amd module, then read the parameter back to confirm. In the cloud the physical host is the provider’s job, but if you run KVM inside a rented VM, that layer is still yours to check.
Where each vendor stands
| Target | Status at time of checking |
|---|---|
| nginx.org | 1.31.9 and 1.30.7 released on 2026-11-18 |
| Official Docker image | nginx:1.31.9, 1.30.7, latest and stable checked on arm64 and all report fixed versions |
| Debian nginx | bookworm 1.22.1, trixie 1.26.3, forky and sid 1.30.6, all marked vulnerable in the tracker |
| Ubuntu nginx | CVE page lists jammy to stonking as needed; newest archive packages date from 26 October; no USN |
| Debian linux | bookworm 6.1.192 and trixie 6.12.117 marked vulnerable; sid 7.3.5 fixed |
| Red Hat | both NGINX CVEs rated Critical, nginx on RHEL 8, 9 and 10 marked Affected with no fix advisory yet; CVE-2026-99312 rated Moderate, kernel on RHEL 9 and 10 marked Fix deferred, RHEL 8 not affected |
| CISA KEV | version 2026.11.18, 1,786 entries, none of the three listed |
The record for CVE-2026-99312 only names fixed releases on the 6.12, 6.18 and 7.3 lines, so check your distribution’s notice for backport status on older stable lines such as 6.1 and 6.6. NGINX derivatives such as OpenResty and Angie, and ingress controllers on Kubernetes, each need checking as well; I did not find a primary notice for every one of them.
If a CDN or WAF sits in front, Wiz says one that terminates HTTP/2 itself may remove some of the attack conditions, depending on the implementation, and that it is no substitute for updating. Whether the origin can be reached directly, bypassing that layer, is still something to test yourself, as with Apache last week.
While you are there, check what these two features are for
While checking the config, it is worth asking whether HTTP/2 and nested virtualisation are still needed. HTTP/2 is often one word added after listen years ago from a tutorial, and nested has simply been on by default since kernel 4.20, so in many environments nobody ever decided to enable either. Only now, when it has to go off, do you find the setting lives in two places, or learn that the switch exists.
Once you know, write it down: which listeners have HTTP/2 on and who needs it, and which hosts really use nested virtualisation. Leave off what nobody uses, and the next time this code has a problem there will be less to check.
Inventory and update
- List every NGINX, including containers, ingress controllers and copies bundled inside products. Use
nginx -vfor the version andnginx -T | grep -n http2to find bothhttp2 on;andlisten ... http2. - Install the fixed package named in your distribution’s or product’s notice and restart, then confirm the running processes are new. The upstream fixed releases for NGINX OSS are 1.31.9 and 1.30.7.
- If no package is available, remove the
http2parameter fromlistenfirst, and on 1.25.1 or later also changehttp2 on;tohttp2 off;. Reload oncenginx -tpasses, then confirm no old worker is leftis shutting down. For listeners carrying gRPC, schedule the update instead, because turning HTTP/2 off would break that traffic. - For HTTPS listeners, test from outside with
curl --http2 --write-out '%{http_version}\n'to print the negotiated version. A result of 1.1 only shows that this listener did not negotiate HTTP/2 this time, so test each virtual host on every listener by its hostname, and do the same at the origin. If an h2c test with--http2-prior-knowledgefails, first rule out that the connection itself is down. - On AMD hosts running KVM, check the kernel version and the nested parameter. Update and reboot where a fixed kernel exists, and turn nested off where nobody uses it.
- After patching, turn HTTP/2 back on for the listeners that need it.
Wiz also suggests watching for unexplained worker crashes or unknown processes started by NGINX. Those are general warning signs, not confirmed indicators of these two flaws.
Notes
- NGINX 1.31.9 and 1.30.7 fix CVE-2026-123007 and CVE-2026-123633. Both need HTTP/2 enabled, and the second also needs proxying with connection slots running short.
- Besides
http2 off;, remove thehttp2parameter fromlisten; in my test, leaving it there kept HTTP/2 on. - After a reload, old workers stay until their connections end, so confirm they have gone. After a package update, restart.
- CVE-2026-99312 only affects AMD KVM hosts with nested on; the fixed kernels are 6.12.118, 6.18.64 and 7.3.5.
- When I checked, Debian and Ubuntu had not fixed nginx, KEV had no entry, and I had found no report of exploitation in the wild.
- The same release also fixes CVE-2026-118899, an infinite loop when reading a request body over TLS 1.3 that does not need HTTP/2, plus issues in HTTP/3, the resolver and the mail proxy, so turning HTTP/2 off is only a stopgap.
Further reading
- nginx security advisories: the nginx.org advisories, a primary source and the reference for affected versions and severity.
- nginx CHANGES: the official change log, a primary source for the security fixes in 1.31.9.
- CVE-2026-123007 (CIRCL), CVE-2026-123633 (CIRCL): mirrors of the CVE records published by F5 as CNA, with CVSS and NGINX Plus versions.
- CVE-2026-99312: the official CVE record, published by the Linux kernel CNA, a primary source.
- Running nested guests with KVM: kernel documentation, a primary source for the nested default and how to set it.
- Wiz: NGINX HTTP/2 vulnerabilities: a write-up by one of the reporters, the source for the mitigation advice and disclosure timeline; it includes exploitation detail.
- Rapid7: CVE-2026-123007: analysis by one of the reporters, the source for the practical attack limits.
- Debian nginx, Debian CVE-2026-99312: Debian’s official tracker, a primary source.
- Ubuntu CVE-2026-123007: Ubuntu’s CVE page, a primary source for per-release status.
- CISA Known Exploited Vulnerabilities Catalog: CISA’s official list, a primary source for vulnerabilities known to be exploited.
- HTTP/2 (Wikipedia), Dangling pointer (Wikipedia): encyclopaedia entries, background on multiplexing, flow control and use-after-free only.
Ideas and technical judgement by Sheng; drafted with Claude · facts researched with Grok, reviewed by Codex; HTTP/2 mitigation tested on OrbStack with nginx:1.31.9, vulnerability not reproduced, KVM part not tested; facts checked up to 08:25 on 19 November 2026 (Taipei time)