NGINX is a web server that can serve files directly or act as a reverse proxy, accepting website and API requests and forwarding them to application servers. A single deployment can sit in front of several services. On November 18, 2026, the NGINX project released Open Source versions 1.30.7 and 1.31.9 to fix two flaws in HTTP/2, a protocol clients use to exchange requests and responses. Wiz researchers Nir Ohfeld and Tomer Schwartz published their findings the same day after reporting both bugs to F5. NGINX overview · Stable release notes · Mainline release notes · Wiz Research disclosure
Affected builds of NGINX Open Source and NGINX Plus can be exposed when NGINX itself handles HTTP/2. Both flaws can crash a worker, one of the NGINX processes that handle requests. In CVE-2026-123007, NGINX can miscount how much data it is allowed to send and write past the memory it set aside. In CVE-2026-123633, a worker running low on room for new connections can release memory that a request it is forwarding still needs. A crash can interrupt requests across the services behind that worker. Canonical says an attacker may also be able to run code remotely if address space layout randomization (ASLR), a defense that makes memory locations harder to predict, is disabled or bypassed. Canonical on CVE-2026-123007 · Canonical on CVE-2026-123633 · NGINX security advisories
Wiz says it ran code remotely without logging in during lab tests and bypassed ASLR. For CVE-2026-123007, it demonstrated code execution on official NGINX 1.31.7 and 1.31.8 builds serving ordinary static files. That matters because NGINX handles the connection itself, before any login in the application behind it. Whether either flaw was used against real deployments, or which other builds and configurations would allow code execution, remains unknown. Wiz Research disclosure
What was disclosed and when
- October 21. Wiz reported the memory-reuse flaw, later assigned CVE-2026-123633, to F5 with lab results and a proposed fix. Wiz says F5 confirmed that another research team had already reported this flaw. Wiz disclosure timeline
- November 10–17. On November 10, F5 supplied a patch for CVE-2026-123633 for review. Wiz reported the second HTTP/2 flaw on November 16. The next day, F5 confirmed that the report matched CVE-2026-123007 and that the planned November 18 release covered it. Wiz disclosure timeline
- November 18. NGINX shipped both fixes. The project credited Xint.io for CVE-2026-123007 and Albert of Vexera AI for CVE-2026-123633. Canonical published separate Ubuntu assessments that day. NGINX release notes · Canonical on CVE-2026-123007 · Canonical on CVE-2026-123633
The upstream NGINX advisories list affected Open Source versions from 1.9.5 through 1.31.8 for CVE-2026-123007 and 1.19.7 through 1.31.8 for CVE-2026-123633. Both ranges exclude 1.30.7 and later 1.30 stable releases. The mainline fix starts at 1.31.9 for both. For exact NGINX Plus release coverage, see the F5 advisories linked from the NGINX security page. NGINX security advisories
Distribution packages can follow different version rules. Canonical identifies CVE-2026-123633 exposure on cleartext HTTP/2 listeners, which run without TLS, in Ubuntu 16.04, 18.04, 20.04, and 22.04. Those packages predate the lower bound in the upstream range. Canonical listed Ubuntu 24.04 as vulnerable to both flaws on November 18, with no fixed package revision in either status table. Check your distributor’s record for both CVEs, including any fix it has applied to an older release. Ubuntu status for CVE-2026-123007 · Ubuntu assessment for CVE-2026-123633
How to check your NGINX exposure and running fix
If you operate NGINX, trace the traffic to each running instance, then verify the build used by its workers. After the update, a health request can show whether a particular replica answered on the path you tested. Use retained logs and host records to investigate possible misuse separately.
Map where HTTP/2 ends
Trace the connection from the client to the content delivery network (CDN) or load balancer, then to each NGINX origin behind it. An HTTP/2 response from the public hostname may only show that the CDN handled HTTP/2. Check which protocol the CDN uses to reach NGINX and compare it with what NGINX records. Include other routes, such as direct public access and private load balancers. An origin receiving HTTP/1.1 from the CDN can still accept HTTP/2 on another path.
Check build support and listener settings separately. The http2 on; directive appeared in NGINX 1.25.1. Older configurations can enable HTTP/2 with the http2 parameter on listen. Check the effective settings for the relevant virtual host and port. If you use nginx -T, keep its output private because it prints the configuration and any sensitive values it contains. NGINX HTTP/2 module reference · NGINX listener reference · How to check NGINX configuration
Verify the build behind each serving worker
nginx -v prints the version of the executable you invoke. nginx -V also shows its build options, which can help you check for HTTP/2 support. For a packaged install, record the distribution release and full package revision, then compare that revision with both vendor advisories. For a container, match the image digest reported for each running replica to the image you verified. The executable on disk can already be updated while an older process still serves requests. NGINX command-line reference
Record the process IDs and start times of the NGINX master, which manages the workers, and the workers that serve requests. On Linux, where you have access, use /proc/<pid>/exe to check the running executable against the verified build. Compare the executable’s file identity or hash. A path alone is insufficient because an update can replace the file at that path while the old process keeps running. Follow your deployment’s upgrade procedure. A configuration reload starts new workers under the existing master, and old workers can keep serving existing clients as they shut down. Confirm that the old workers have exited and that each replica is serving with the expected build. NGINX process control · Linux process executable reference
In this synthetic worksheet, the CDN ends the client’s HTTP/2 connection and reaches both origins over HTTP/1.1. A separate internal load balancer passes the encrypted connection through to NGINX. NGINX then handles HTTP/2 itself. The package revision, process IDs, and times are illustrative.
| Evidence | proxy-a |
proxy-b |
|---|---|---|
| Client to CDN / CDN to origin | HTTP/2 / HTTP/1.1 | HTTP/2 / HTTP/1.1 |
| Direct public access / internal path | Blocked / encrypted connection passed through | Blocked / encrypted connection passed through |
| NGINX listener | listen 443 ssl; with http2 on; |
listen 443 ssl http2; |
| Package or build | Upstream 1.30.7 in a container image | Ubuntu 24.04, example revision 1.24.0-2ubuntu7.5 |
| Running executable | Image digest and worker executable match the verified 1.30.7 image | Worker executable matches the installed Ubuntu package |
| Master / worker | 4100 / 4107, replaced at 17:10 UTC, old workers gone |
5200 / 5208, restarted at 17:13 UTC, old workers gone |
| Patch status | Fixes verified for both CVEs | Unresolved |
proxy-b still needs a build with verified fixes for both CVEs. If a fix is pending, verify that any temporary mitigation covers the internal path. The CDN’s use of HTTP/1.1 leaves that path open.
Check that each replica answers a health request
From an authorized network, send a normal GET to an existing read-only health endpoint. Use an approved route that targets one replica at a time and a hostname that matches its certificate. If the endpoint accepts a query marker and your access logs retain it, use a unique, non-sensitive value. Substitute your hostname and marker in the example, then repeat for the second replica. An HTTP/2-capable curl build already attempts HTTP/2 negotiation over HTTPS by default. The server can still negotiate HTTP/1.1, so read the protocol in the output along with the status. curl command reference
curl --silent --show-error --max-time 5 --output /dev/null \
--write-out 'status=%{http_code} protocol=%{http_version}\n' \
'https://proxy-a.ops.example/healthz?check=rollout-a-1711'
Match the marker to an origin access row and verify the replica identity. For this check, look for HTTP/2 and the status your health endpoint is supposed to return. A cached CDN response may never reach NGINX. A request through a load balancer can keep hitting the same replica.
These synthetic startup and access rows use a normalized format. The fields available in your logs depend on how you collect them.
2026-11-18T17:10:02Z proxy-a startup master=4100 worker=4107
2026-11-18T17:11:04Z proxy-a access worker=4107 "GET /healthz?check=rollout-a-1711 HTTP/2.0" 200
2026-11-18T17:13:08Z proxy-b startup master=5200 worker=5208
2026-11-18T17:14:11Z proxy-b access worker=5208 "GET /healthz?check=rollout-b-1714 HTTP/2.0" 200
If you already collect these logs in Fluxtail, open the NGINX stream and narrow the time window to the rollout, such as 17:09–17:16 UTC. Filter by service and replica host if your collector maps those fields. Search for rollout-a-1711 or rollout-b-1714 and open the matching raw row to check the timestamp, status, and protocol. Clear the marker search and keep the time and replica filters to read nearby startup and error rows. New probe rows can appear as they arrive in Live Tail.
In the example, each replica answered one HTTP/2 request at the time shown. The first time each replica recovered and the health of other routes remain unknown. The running-build check is what establishes the patch status in the worksheet.
Investigate suspicious activity separately
Review retained NGINX error and service logs for unexpected worker exits, repeated restarts, or connection failures during the time each affected listener was reachable. Compare those times with deployment records and available host or container audit events. An unexpected process started by an NGINX worker or an unexplained outbound connection gives you a lead to check. A crash has many possible causes, and a connection that fails early may leave no access row. Record which replicas and time ranges your logs cover. An empty search cannot rule out compromise. How to read logs during an incident