Skip to content
Zyberum Cyber Security Firm
Menu
VulnerabilitiesWeb

NGINX2Shell (CVE-2026-123007): nginx HTTP/2 Flaws, Detection, Fix

nginx 1.31.9 and 1.30.7 fix two HTTP/2 memory bugs and eight further issues. Which setups are exposed, how to check them and why a reload is not enough.

Zyberum Security Team · Published · 17 min read

On 18 November 2026 nginx 1.31.9 (mainline) and 1.30.7 (stable) were released with ten security fixes. Two of them are in the HTTP/2 module: CVE-2026-123007, a memory corruption in flow control, and CVE-2026-123633, a use after free. Both are reachable without authentication where HTTP/2 is enabled, the second one only under additional conditions. The usual result is a crashed worker process. According to F5, code execution is possible where ASLR is disabled or can be bypassed.

The names NGINX2Shell and NGINX2Root are not vendor terms. Neither nginx.org nor the text of the CVE records uses them. “Nginx2Shell” is used by JFrog and others for the two HTTP/2 bugs, “NGINX2Root” is the name Xint gave its set of five nginx bugs plus one Linux kernel bug, and Red Hat adopted it as the title of its bulletin. As of 19 November 2026 no exploitation in the wild has been reported, but a proof of concept for CVE-2026-123007 is public. This guide is part of our series on the 2Shell vulnerabilities.

The facts in one table

All ten issues are fixed in nginx 1.31.9 and 1.30.7. The one exception in the table is the PROXY protocol bug, which exists only in the 1.31 mainline branch and is fixed in 1.31.9.

CVEModuleBug classNeeds in the configurationAffectedSeverity, by whom
CVE-2026-123007HTTP/2 (ngx_http_v2_module), flow controlInteger underflow leading to a heap out-of-bounds write (CWE-191, CWE-787)HTTP/2 enabled, nothing else1.9.5 to 1.31.8nginx.org: major. F5: CVSS 4.0 9.2, CVSS 3.1 8.1. Red Hat: Critical. Ubuntu: High
CVE-2026-123633HTTP/2 (ngx_http_v2_module)Use after free (CWE-416)HTTP/2 enabled, proxying to a backend, worker_connections running short1.19.7 to 1.31.8nginx.org: medium. F5: 9.2 and 8.1. Red Hat: Critical
CVE-2026-123634ResolverUse after free (CWE-416)resolver, a proxied server name with variables, an SSI or mirror subrequest that resolves the same name0.6.18 to 1.31.8nginx.org: medium. F5: 9.2 and 8.1. Red Hat: Important
CVE-2026-121810HTTP/3 (ngx_http_v3_module)Use after free (CWE-416)HTTP/3 enabled (listen ... quic), worker_connections running short1.25.0 to 1.31.8nginx.org: medium. F5: 9.2 and 8.1. Red Hat: Critical
CVE-2026-113425Mail proxy (ngx_mail_proxy_module)Use after free (CWE-416)A mail {} block that acts as a proxy0.5.33 to 1.31.8nginx.org: major. F5: 9.2 and 8.1. Red Hat: Critical
CVE-2026-118899Request body over TLSInfinite loopReading a client request body over TLS 1.31.13.0 to 1.31.8nginx.org: major
CVE-2026-120993PROXY protocolNULL pointer dereferencePROXY protocol v21.31.4 to 1.31.8nginx.org: minor
CVE-2026-125283alias with try_filesBuffer overflowalias used together with try_files0.9.6 to 1.31.8nginx.org: minor
CVE-2026-120967Stream, TLSSNI bypass with client certificate validationstream with client certificate validation1.25.5 to 1.31.8nginx.org: medium
CVE-2026-123640Stream, ssl_prereadInsufficient validationssl_preread in stream1.15.2 to 1.31.8nginx.org: minor

The ratings do not agree. F5, which assigned the CVE ids, scores all of the first five at 9.2 on CVSS 4.0 and 8.1 on CVSS 3.1 with the vector CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H. nginx.org labels three of the same five as “medium”. NVD had not published its own score on 19 November. For the last five rows we list the nginx.org label only.

The exploitation status is also described differently. The CISA entry on the CVE records says “Exploitation: none” although Rapid7 has published a proof of concept for CVE-2026-123007. None of the CVEs is in the CISA KEV catalogue. Wiz and Xint state that they have working exploits and do not release them.

Which installations are exposed?

For the two HTTP/2 bugs: every nginx from the affected version range that accepts HTTP/2 connections from a client an attacker can control. nginx speaks HTTP/2 only where the configuration enables it, but most builds contain the module and many configurations switch it on.

CVE-2026-123007: the conditions, and what http2 off removes
affected version1.9.5 to 1.31.8HTTP/2 modulecompiled inHTTP/2 enabledlisten or http2 onclient connectsALPN h2 negotiatedHTTP/2 enabledlisten or http2 onclient connectsALPN h2 negotiatedhttp2 off
  1. Step 1/5The version is in the affected range: 1.9.5 to 1.31.8 for CVE-2026-123007, 1.19.7 to 1.31.8 for CVE-2026-123633. A distribution package with a backported fix is the exception.
  2. Step 2/5The build contains the HTTP/2 module. Most builds do: nginx -V shows --with-http_v2_module.
  3. Step 3/5The configuration enables HTTP/2, either with listen ... http2 or with http2 on; (nginx 1.25.1 and later).
  4. Step 4/5A client can open an HTTP/2 connection to this nginx. For CVE-2026-123007 nothing more is needed: it was confirmed on a server that only delivers static files.
  5. Step 5/5Interim measure: with http2 off; and no http2 on the listen lines, clients fall back to HTTP/1.1 and the affected code is not reached. This comes from researchers and Red Hat, F5 lists no workaround.

The other three memory bugs have their own switches:

CVEDirectives that enable the affected code
CVE-2026-123007listen ... http2 (syntax of 1.9.5 to 1.25.0, still accepted later) or http2 on;
CVE-2026-123633HTTP/2 as above and a backend: proxy_pass, fastcgi_pass, uwsgi_pass, scgi_pass, grpc_pass and similar, while free worker connections run low
CVE-2026-123634resolver, a proxied server name built from variables, and an SSI or mirror subrequest
CVE-2026-121810listen ... quic, that is HTTP/3. Disabling HTTP/2 does not help here
CVE-2026-113425A mail {} block that proxies, according to Xint with proxy on and auth_http

What a CDN or load balancer in front changes

It reduces the exposure of the HTTP/2 bugs and does not remove it. What counts is which system terminates the HTTP/2 connection of the client.

Where HTTP/2 ends decides which hop is exposed
Clientor attackerCDN / LBterminates HTTP/2nginxorigin or ingressUpstreamapplicationHTTP/2 directclient frames reach nginxHTTP/2ends at the CDNnew connectionHTTP/1.1 or HTTP/2 from the CDNorigin bypassorigin IP reachable directlyproxy_passneeded for CVE-2026-123633allow only the CDNfirewall the origin
  1. Step 1/6Setup A: nginx is reachable from the internet and terminates HTTP/2 itself. The frames of the client reach the affected code directly. JFrog names these as most exposed: internet-facing servers, ingress controllers, load balancers and appliances.
  2. Step 2/6Setup B: a CDN or load balancer in front terminates the HTTP/2 connection of the client.
  3. Step 3/6The CDN opens its own connection to nginx. Rapid7 expects that the frames of the attacker are then not passed on in the form the bug needs. Wiz says this may mitigate parts, depending on the implementation.
  4. Step 4/6This holds only as long as nobody reaches the origin directly. An origin address that answers from the internet, or a client inside the network, is setup A again.
  5. Step 5/6The hop from nginx to the backend matters for CVE-2026-123633: that bug needs HTTP/2 from the client and proxying to an upstream. CVE-2026-123007 does not need a backend.
  6. Step 6/6Restrict the origin to the addresses of the CDN or load balancer, test the origin itself and not the CDN name, and update anyway.

A secondary source reports statements by CDN vendors that origins reachable only through their proxy are protected while you update. We could not confirm these statements at the vendors and treat them as unconfirmed.

Packages and derivatives on 19 November 2026

The status below is a snapshot and will change. “Not yet fixed” means that no fixed package was listed when the page was read.

ProductStatusDetail
nginx open sourceFixed1.31.9 and 1.30.7. Legacy branches such as 1.28 and 1.26 received no upstream release
NGINX PlusFixed37.3.0.1 (CR) and 37.0.8.1 (LTS) cover both HTTP/2 bugs. CVE-2026-123633 alone was already fixed in 37.2.0.1 and 37.0.7.1 of 14 October 2026. For older release trains the CVE records name no fixed build
Official Docker image nginxFixedTags 1.31.9, mainline, latest, 1.30.7, stable and their variants. Older pinned tags such as stable-bookworm were not rebuilt
UbuntuFixed, USN-9159-126.10: 1.30.4-7ubuntu1.2, 26.04: 1.28.3-2ubuntu1.13, 24.04: 1.24.0-2ubuntu7.20, 22.04: 1.18.0-6ubuntu14.23. 20.04, 18.04 and 16.04 only with Ubuntu Pro
DebianPartly fixedtrixie: 1.26.3-3+deb13u11 (DSA-6642-1), sid: 1.30.6-4. bookworm (1.22.1-9+deb12u10) and forky (1.30.6-3) not yet fixed
AlpineFixededge and 3.24: 1.30.7-r0, 3.23 and 3.22: 1.28.3-r8
RHEL 8, 9, 10Not yet fixedAffected, no errata listed. Red Hat states that it is expediting the fixes. The 3scale image apicast-gateway-rhel8 is affected as well
nginx.org RPM repositoryFixed, with a catchThe 1.30.7 packages carry no security metadata, so dnf upgrade --security does not select them. Run a plain upgrade
OpenRestyNot yet fixedThe latest release 1.31.8.1 bundles nginx 1.31.8. An issue is open, without an answer from the maintainers
Angie, freenginx, TengineNot statedNone of the projects had published a statement or a release that names these CVEs. The latest freenginx release is 1.31.5, the latest Tengine release builds on the nginx 1.31.5 core. Treat them as affected until the project says otherwise
Kubernetes ingress-nginx (community)Will not be fixedThe repository is archived since 24 March 2026 and receives no security updates. The last release v1.15.1 bundles nginx 1.27.1, with HTTP/2 enabled by default on port 443
F5 NGINX Ingress Controller, NGINX Gateway FabricUnclearNew releases appeared on 18 November (v5.8.0, v2.9.0). The release notes do not name the bundled nginx version. Check the version in the running image
AlmaLinux, Amazon Linux, SUSEReported, unconfirmedA secondary source reports fixed packages for AlmaLinux and Amazon Linux and an ongoing analysis at SUSE. Check the advisory of the distribution

Products that bundle their own nginx, such as API gateways, NAS systems and appliances, are not patched by an update of the host operating system. JFrog states that the nginx bundled with Artifactory normally has HTTP/2 off and that a patched release is in progress. For other vendors, ask for a statement. One user report describes the AKS app routing add-on as shipping nginx 1.27.1 with HTTP/2 on and without a supported switch to disable it. The maintainers had not answered when we checked.

How do the HTTP/2 bugs work?

Both are memory-safety bugs in ngx_http_v2_module: one is a counter that goes below zero and then permits a write past a buffer, the other is a connection that is released while a request still uses it.

CVE-2026-123007: flow control

HTTP/2 uses flow control: each side keeps a window, a number that says how much data may still be transferred. The CVE record describes an integer underflow in the accounting of this window. After the underflow the window “can exceed its allowed threshold”, and nginx writes beyond a buffer on the heap. The nginx changelog calls it “a flow control window overflow in HTTP/2”.

The figure shows the principle. It is schematic: the advisories publish neither buffer sizes nor the memory layout.

An accounting value below zero, a write past the buffer
HEAP OF ONE NGINX WORKER (SCHEMATIC)other databufferfixed sizeneighbouring heap dataother objects of the workerwindow is positive: write stays insideaccounting value drops below zerowindow now above its allowed limitwrite runs past the end of the bufferworker exits on signal 11
  1. Step 1/5In normal operation the window is a positive number. nginx copies data into a buffer, and the window keeps the amount inside the buffer.
  2. Step 2/5A sequence of HTTP/2 frames makes the accounting subtract more than the window holds. The value goes below zero (integer underflow, CWE-191).
  3. Step 3/5After the underflow the window is no longer a small number. In the words of the CVE record it can exceed its allowed threshold, so more data counts as permitted than the buffer can take.
  4. Step 4/5The next write runs past the end of the buffer and overwrites neighbouring data on the heap of the worker (out-of-bounds write, CWE-787).
  5. Step 5/5Normally the worker then ends with a segmentation fault. F5 states that code execution is possible where ASLR is disabled or can be bypassed.

The bug affects the data plane only and needs no special configuration beyond HTTP/2. It was confirmed over HTTPS on 64-bit systems.

CVE-2026-123633: use after free under connection pressure

The second bug needs more conditions. nginx accepts HTTP/2 from clients and proxies the requests to a backend. When the worker runs short of free worker_connections, nginx may reuse or close a connection that still serves a request. The memory of that connection is then used after it was freed. The changelog describes it as a use after free “when using HTTP/2 if worker_connections were not enough”.

CVE-2026-121810 is the same family on HTTP/3: a use after free when worker_connections are not enough, according to Xint reached when HTTP/3 requests are forwarded to an upstream.

How hard is exploitation?

The sources disagree, and the honest answer is: a crash is the expected outcome, code execution has been shown by researchers under their conditions.

  • F5 ties code execution to ASLR being disabled or bypassed. Wiz writes that ASLR does not protect against the two HTTP/2 issues and that its lab exploits include a bypass. Xint reports working exploits against endpoints in public clouds.
  • Rapid7 describes the limits of its published proof of concept for CVE-2026-123007: about 128 GB of HTTP/2 data within the default connection lifetime of one hour, and a dependency on one specific memcpy variant of glibc on x86-64. That describes this proof of concept and is not a safe threshold.
  • Exposure numbers in the press are counts of version banners, not of confirmed vulnerable servers, and the published totals differ between sources.

How to identify it

Check four things per instance: the version of the package, whether the configuration enables the affected modules, whether HTTP/2 is negotiated from outside, and whether the process that runs is the patched binary.

1. Version and build

nginx -v 2>&1
nginx -V 2>&1 | grep -- '--with-http_v2_module'

# Packages: compare with the tracker of your distribution
dpkg -l 'nginx*' | awk '/^ii/{print $2, $3}'
rpm -q nginx
apk info -v nginx

Upstream versions 1.9.5 to 1.30.6 and 1.31.0 to 1.31.8 are affected by CVE-2026-123007. Distribution packages keep the old upstream number after a backport, so nginx -v and the Server banner say nothing about a patched Debian or Ubuntu package. Compare the full package version with the table above or with the tracker of the distribution.

2. Configuration

# HTTP/2 enabled?
nginx -T 2>&1 | sed 's/#.*//' | grep -nE 'listen[[:space:]].*http2|^[[:space:]]*http2[[:space:]]+on[[:space:]]*;'

# Backend present (CVE-2026-123633)?
nginx -T 2>/dev/null | grep -nE '\b(proxy|fastcgi|uwsgi|scgi|grpc|memcached|tunnel)_pass\b'

# HTTP/3, resolver, mirror, SSI, mail proxy (the other fixes)
nginx -T 2>/dev/null | grep -nE 'quic|http3|resolver|mirror|ssi[[:space:]]+on|^[[:space:]]*mail[[:space:]]*\{'

nginx -T prints the configuration on disk. Until the next reload it can differ from what the running process uses.

3. Check from outside whether HTTP/2 is negotiated

openssl s_client -connect HOST:443 -alpn h2 </dev/null 2>/dev/null | grep ALPN
curl -sI --http2 https://HOST -o /dev/null -w '%{http_version}\n'

# On the host: HTTP/3 and mail listeners
ss -lunp 'sport = :443'
ss -ltnp | grep -E ':(25|110|143|465|587|993|995)'

ALPN protocol: h2 or the output 2 means that HTTP/2 is negotiated. Run the check against the origin address and not only against the public name: a test through the CDN tests the CDN.

4. Containers and Kubernetes ingress controllers

docker ps --format '{{.Image}}'
kubectl get pods -A -o jsonpath='{range .items[*]}{.metadata.namespace}{"\t"}{.spec.containers[*].image}{"\n"}{end}' | grep -iE 'nginx|openresty|kong|apisix'
kubectl exec -n NS POD -- nginx -v

Read the version inside the container, not from the tag. Images that bundle nginx need a rebuild on a fixed base image and a new deployment: restarting a running container keeps the old image.

Every release of the community ingress-nginx controller bundles an affected nginx, with HTTP/2 enabled by default on port 443. Since the project is retired, this finding does not go away with an update.

5. Verify that the new binary is running

A reload is not enough. nginx -s reload re-reads the configuration and starts new workers from the same master process. It does not start a new executable. The nginx documentation describes two ways to change the binary: the upgrade sequence with the signals USR2, WINCH and QUIT, or a full restart.

The upgrade is not finished until the old process is gone
package updatenew file on diskreloadsame masterold binarystill in memoryfull restartor USR2 upgradeverify/proc/PID/exe
  1. Step 1/5The package manager replaces the nginx executable on disk. The running master process still executes the old file.
  2. Step 2/5A reload starts new workers, but from the same master. The code in memory does not change.
  3. Step 3/5The result looks healthy and is still vulnerable: new package version, old binary in every process.
  4. Step 4/5A full restart, or the binary upgrade with USR2, WINCH and QUIT, starts the new executable. A restart disconnects every client.
  5. Step 5/5Verify: no process may point to a deleted executable, and no worker may remain in the state "shutting down".
systemctl restart nginx

# "(deleted)" behind the path means: this process still runs the replaced file
for p in $(pgrep -x nginx); do echo "$p $(readlink /proc/$p/exe)"; done

# Start time of master and workers, and workers that are still shutting down
ps -eo pid,ppid,lstart,cmd | grep '[n]ginx'

Workers shown as “shutting down” run the old code until their last connection ends. The Ubuntu notice says only that a standard system update makes all necessary changes and does not mention a restart. Verify it yourself.

6. Log and crash signs

All sources call the following general warning signs, not confirmed indicators. JFrog writes that there are no confirmed indicators of compromise yet.

grep -Ei 'worker process.*exited|signal 11|segfault' /var/log/nginx/error.log
coredumpctl list nginx
docker logs C 2>&1 | grep "exited on signal"
ss -tinp state established '( sport = :443 )'
WhereSign
error.logworker process N exited on signal 11, in series
error.logN worker_connections are not enough, reusing connections, followed by crashes (CVE-2026-123633, CVE-2026-121810)
error.logcould not cancel <name> resolving alerts, especially for unfamiliar names (resolver bug)
ConnectionsLong-lived HTTP/2 connections from one address with very large inbound and small outbound byte counts, often together with many short TLS connections from the same address
Access logLittle. The relevant frames, such as WINDOW_UPDATE and RST_STREAM, are not requests. A log_format with $connection, $connection_requests and $request_length helps to compare received bytes with logged requests
Process auditAny program started by the user of the nginx workers. Record execve and execveat with auditd and investigate at once

Look back at least one week. One source reports that the HTTP/2 fixes were visible in the public repository from 11 November, a date we could not confirm.

Do not expect a network signature to help. None of the sources names an IDS rule, a WAF rule or a scanner plugin. The traffic is inside TLS and at the framing layer of HTTP/2.

What to do

Update, restart, verify. The update is the only complete fix, and F5 lists no workaround for the two HTTP/2 bugs.

  1. Update nginx to 1.31.9 or 1.30.7, NGINX Plus to 37.3.0.1 (CR) or 37.0.8.1 (LTS), or install the fixed package of your distribution. Legacy upstream branches get no fix: rely on a distribution backport or move to a supported branch.
  2. Restart or upgrade the binary, then verify the running process as in step 5 above. A reload leaves the old executable in memory.
  3. Order by exposure: internet-facing HTTP/2 endpoints first, then reverse proxies, then the rest.
  4. Rebuild and redeploy container images on a fixed base. Pinned older tags of the official image were not rebuilt.
  5. Plan the migration off ingress-nginx. The community controller will not receive a fix. Disabling HTTP/2 there is a stopgap, not a state to stay in.
  6. Where no fixed package exists yet (Debian bookworm and forky, RHEL on 19 November): apply the interim measures below or use the nginx.org packages. On RPM systems with the nginx.org repository run a plain upgrade, because the --security filter misses the package.
  7. Ask vendors of bundled copies for their status: OpenResty-based products, API gateways, NAS systems, appliances.
  8. Review logs for the signs above. Where a worker of nginx has started another program, treat the host as compromised.

If you cannot update today

These measures remove the precondition of single bugs. They are temporary, and each covers only the CVEs named in its row.

MeasureEffectLimit
nginx 1.25.1 and later: http2 off; in every affected server block and remove http2 from the listen lines. Older versions: change listen 443 ssl http2; to listen 443 ssl;Closes the path to CVE-2026-123007 and CVE-2026-123633Clients fall back to HTTP/1.1. Native gRPC and everything that needs HTTP/2 end to end breaks. Not a workaround from F5, recommended by researchers and Red Hat
No SSI or mirror, no proxied server name with variables. Point proxy_pass at a fixed upstream groupWorkaround from F5 for CVE-2026-123634, the fixed upstream group is the advice of XintSwitching off mirror alone may not be enough, other subrequest features may reach the code
Remove quic from listen, set http3 off;, send add_header Alt-Svc 'clear' always;Closes the path to CVE-2026-121810HTTP/3 clients fall back to TCP
Restrict mail listeners to trusted networks or bind them to loopback, or remove the mail contextFor CVE-2026-113425. F5 names the restriction to trusted networks, Xint the loopback binding, Red Hat the removalRed Hat calls the restriction to trusted clients a partial mitigation
ingress-nginx: use-http2: "false" in the ConfigMapReported as a stopgap for the HTTP/2 bugs, not confirmed by a primary sourceNot possible on managed add-ons that forbid ConfigMap edits. The controller stays unpatched
Firewall the origin so that only the CDN or load balancer reaches itThe CDN terminates the HTTP/2 connection of the clientDepends on the implementation of the CDN, does not cover internal clients

Apply and check the HTTP/2 change like this:

nginx -t && nginx -s reload
openssl s_client -connect HOST:443 -alpn h2 </dev/null 2>/dev/null | grep ALPN

After the reload the existing connections stay on the old workers until they close. The measure is effective once those workers have exited and ALPN no longer offers h2.

Hardening that limits the damage

None of this prevents the memory corruption. It decides how far an attacker gets afterwards and how fast you notice.

  • Keep ASLR enabled, and do not rely on it. F5 conditions code execution on ASLR being disabled or bypassed, Wiz states that it does not protect against the HTTP/2 issues.
  • Block the SCTP kernel module where it is not needed. The chain that Xint describes from nginx worker to root uses a separate bug in the Linux kernel SCTP code, which the nginx update does not fix. On hosts without a need for SCTP, a line install sctp /bin/false under /etc/modprobe.d/ prevents loading the module. It does not unload a loaded module (lsmod | grep sctp) and has no effect where SCTP is built into the kernel. We found neither a CVE id nor a patch status for this kernel bug.
  • Restrict origins to the CDN or load balancer in front of them.
  • Alert on worker crashes and on processes started by the nginx user.
  • Do not run “proof of concept” repositories from unknown authors on production systems. Sources warn of fake repositories.

Two earlier nginx advisories in 2026

The release continues a series: nginx.org lists 36 advisories with CVE ids from 2026. Two earlier ones follow the same pattern of a heap overflow behind a specific configuration, with the same F5 scores of 9.2 (CVSS 4.0) and 8.1 (CVSS 3.1).

CVE-2026-42945CVE-2026-42533
Modulengx_http_rewrite_modulemap with regular expressions
BugHeap buffer overflowHeap buffer overflow in a worker
ConfigurationA rewrite with an unnamed capture ($1, $2) and a ? in the replacement, followed by another rewrite, if or setA map with regex matching, where a string expression references the capture variables before the output variable of the map
Published13 May 202615 July 2026
nginx.org labelmediummajor
Affected0.6.27 to 1.30.00.9.6 to 1.31.2
Fixed1.31.0, 1.30.1. NGINX Plus R36 P4, R32 P61.31.3, 1.30.4. NGINX Plus 37.0.3.1, R36 P7
ExploitationVulnCheck reported exploitation attempts against its canaries within days, according to SecurityWeek. Not in CISA KEVThe CISA entry records a proof of concept. Not in CISA KEV

CVE-2026-42945 is also known as “NGINX Rift”, a name from the finder depthfirst and again not from nginx.org. The finder’s interim measure was to replace unnamed captures with named captures. Both are fixed in every version that contains the November fixes, so one update to 1.31.9 or 1.30.7 closes all three rounds.

What it means for a penetration test

A reverse proxy is part of the attack surface of every application behind it. In a web application penetration test we record the front end as well: which nginx build answers, which protocols it negotiates, whether the origin can be reached past the CDN, which ingress controller runs in the cluster. For these vulnerabilities we verify exposure through versions and configuration. We do not send triggering traffic to production systems, because the first visible effect of a hit is a crashed worker and dropped connections.

FAQ

Frequently asked questions

Is a reload enough after the package update?

No. A reload re-reads the configuration and starts new workers from the same master process, so the old executable keeps running. Restart nginx or use the binary upgrade sequence described in the nginx documentation, then check that master and workers run the new file.

We do not use HTTP/2. Do we still have to update?

Yes. Without HTTP/2 the two HTTP/2 bugs are not reachable, but the same release fixes eight further issues, among them use-after-free bugs in the resolver, in HTTP/3 and in the mail proxy. Which of them apply depends on your configuration, and the update covers all of them.

Does a CDN or load balancer in front of nginx protect us?

Only partly. If the CDN terminates HTTP/2, nginx usually receives the connection of the CDN and not the frames of the attacker. That holds only if the origin cannot be reached directly, and it does not cover internal clients. None of the researchers presents it as a fix.

nginx -v shows an old version on Debian or Ubuntu. Are we vulnerable?

Not necessarily. Distributions backport the fix and keep the old upstream number, for example 1.24.0-2ubuntu7.20 on Ubuntu 24.04. Compare the full package version with the security tracker of your distribution, not with the upstream number.

Call usBook a call

Pick a time that suits you

Open in a new tab