Red Hat scores CVE-2026-119618 at 9.8 and calls the attack complexity low. The CVE record scores the same bug at 8.1 and calls the attack complexity high. The Apache Software Foundation rates it critical and publishes no number at all. One flaw, three verdicts, and the distance between them decides whether your change window is tonight or Thursday.
The flaw itself is narrow and specific. Apache HTTP Server 2.4.71 shipped on November 11 with a fix for a concurrency bug in mod_http2. The project advisory describes it as a "Concurrent Execution using Shared Resource with Improper Synchronization vulnerability in Apache HTTP Server's mod_http2 in assign_headers()" that can lead to remote code execution. Affected releases run from 2.4.55 through 2.4.70, plus the standalone mod_h2 module from GitHub at 2.0.8 through 2.0.42. Anthropic's OSS Scanner reported it to the security team on October 26 as ANT-2026-MAPKT6MZ, and Anthony Weems and Simon Scannell of Google Cloud Vulnerability Research reported it on November 3. The fix landed as commit 824beab. The same release also closes CVE-2026-106023, a NULL pointer dereference in mod_md rated important and reaching back to 2.4.30, and CVE-2026-109163, a one-byte stack overwrite in mod_vhost_alias rated low.
Three conditions have to hold together before any of this is your problem: mod_http2 has to be loaded, h2 or h2c has to appear in an active Protocols directive, and the server has to be reachable by the client that sends the requests. Red Hat states that Red Hat Enterprise Linux 8 and earlier are not affected. Canonical lists Ubuntu 18.04 and older as not affected. If your edge terminates HTTP/2 at a CDN or a load balancer and speaks HTTP/1.1 to the origin, the origin is not reachable along this path, whatever version string it reports. Where none of the three conditions holds, leave this in the normal patch cycle and spend the morning elsewhere. As of this morning the CVE is absent from the CISA Known Exploited Vulnerabilities catalog, which still stands at catalog version 2026.11.13 and 1,781 entries, and no vendor has reported exploitation in the wild.
What assign_headers gets wrong
mod_http2 serves concurrent streams on worker threads while the connection itself is owned by another thread. Red Hat's security bulletin RHSB-2026-017 describes the failure precisely: a worker thread writes into memory owned by the connection thread through a shared allocator that lacks adequate locking. A crafted sequence of concurrent HTTP/2 requests from an unauthenticated client corrupts the heap of an httpd child process. No credentials, no prior access, no user interaction.
The interesting part is what Red Hat says next. Its CVE page carries a 9.8 with the vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, and then rates RHEL 9 and RHEL 10 as important rather than critical. The reason given is structural: exploitation depends on HTTP/2 stream workers sharing an allocator with the connection, and the packages in those two releases allocate each stream from a separate allocator, which "substantially increases the complexity of exploiting the flaw." The CVE record's 8.1 reaches the same conclusion through the vector, marking attack complexity high where Red Hat marks it low. Canonical assigns the Ubuntu priority High and publishes the 8.1.
Public research so far matches the more conservative reading. The available proof-of-concept work reproduces memory corruption on 2.4.70 under sanitizers and stops there, with no demonstrated code execution. Where code execution did land, it would land as the account the worker runs under, apache or www-data, and reaching root would need a second bug. That is not a reason to skip the patch. Pre-authentication heap corruption in the process that answers every request is a reliable crash at minimum, and a crash loop against a single-server web stack is an outage an attacker can order on demand.
Three conditions, and the version number answers none of them
Start with the question your scanner cannot ask. Run this against each internet-facing Apache host, from the host and from outside it.
# 1. Is the module even loaded?
apachectl -M 2>/dev/null | grep -i http2 # Debian/Ubuntu: apache2ctl -M
httpd -V 2>/dev/null | head -2 # server version and MPM
# 2. Is h2 or h2c actually enabled, in any vhost?
grep -rniE '^[[:space:]]*Protocols' /etc/httpd/conf/ /etc/httpd/conf.d/ \
/etc/apache2/ 2>/dev/null
grep -rl 'http2' /etc/httpd/conf.modules.d/ /etc/apache2/mods-enabled/ 2>/dev/null
# 3. Does the server negotiate HTTP/2 on the wire? This is the answer that counts.
curl -ksI --http2 https://host.example/ | head -1 # "HTTP/2 200" means enabled
openssl s_client -alpn h2 -connect host.example:443 </dev/null 2>/dev/null \
| grep -i 'ALPN protocol'
curl -ksI --http2-prior-knowledge http://host.example/ | head -1 # plaintext h2c
Condition two trips people up more than the others, because Protocols is valid inside a virtual host. A clean global config with one legacy vhost that enables h2 is an exposed server. The ALPN check settles it: a host that will not negotiate h2 or answer h2c is not reachable through this bug today, regardless of what is installed. Run the plaintext variant too, since h2c on an internal listener is easy to forget.
The answers that change your plan
- Negotiates
h2, package below the distribution's fixed build. Patch it in today's window, or droph2until you can. - Negotiates
h2, package at or above the fixed build. Done, once you confirm the service restarted after the upgrade. - Answers HTTP/1.1 only, module still loaded. Not exposed through this bug. Patch on the normal cycle and note the host, because one
Protocolsedit puts it back in scope. - HTTP/2 terminated at a CDN or load balancer. Check the terminator's own stack and leave the origin alone.
Why the httpd version lies in both directions
The affected component is mod_http2, and distributions ship that module on a schedule of their own. Canonical's tracker entry is the clearest illustration available: Ubuntu 22.04 is listed as fixed at apache2 2.4.52-1ubuntu4.27. Read that against Apache's advisory and 2.4.52 sits below the 2.4.55 floor, which would make the release unaffected. Canonical's own note explains the contradiction, recording that newer module versions were backported into its stable releases. The httpd version told you nothing in either direction, and 24.04 is fixed at 2.4.58-1ubuntu8.19 for the same reason.
Debian runs the other way. The security tracker marks bookworm vulnerable at 2.4.68-1~deb12u1 and fixed at 2.4.71-1~deb12u1 from the security suite, with trixie following the same pattern through DSA-6620-1. Two hosts can both report 2.4.68 while one is patched and one is not. Treat a banner-derived verdict as a hint and get the real answer from the package.
# Debian/Ubuntu: the package version, and the changelog that names the CVE
dpkg-query -W -f='${Package} ${Version}\n' 'apache2*' 'libapache2-mod-http2*' 2>/dev/null
zgrep -il 'CVE-2026-119618' /usr/share/doc/apache2*/changelog.Debian.gz 2>/dev/null
# RHEL family: mod_http2 is its own package on 9 and 10
rpm -q httpd mod_http2 2>/dev/null
rpm -q --changelog mod_http2 2>/dev/null | grep -m2 'CVE-2026-119618'
# Container images keep whatever they were built with
docker ps --format '{{.Image}}' | sort -u
docker run --rm --entrypoint httpd <image> -V | head -2
Scale is worth a sentence, because the circulating numbers are softer than they look. Censys measured 166,240,610 Apache HTTP Server web properties on November 12 and roughly 1.5 million hosts reporting a release in the affected range. Two caveats come from Censys itself: scan data resolves a version on only about 40 percent of Apache hosts, and the platform has no field for HTTP/2 negotiation, so that 1.5 million figure speaks to one of the three conditions and guesses at none of the others. Around 40 percent of the affected hosts sit in five networks, Amazon, Google, Hetzner, OVH and DigitalOcean, and 2.4.71 was appearing on roughly 2,000 hosts a day. Vendor-patched builds that still report an affected version are counted as affected, which is the same trap one layer out.
The stopgap, and what it costs
Where the patch cannot land inside the window, Red Hat's bulletin gives the configuration change: remove h2 and h2c from every Protocols directive so the line reads Protocols http/1.1, then restart with systemctl restart httpd. Verify with curl -kI --http2 https://localhost and treat an HTTP/1.1 200 OK first line as the proof, not the edited file. The cost is real but small for most sites: HTTP/1.1 gives up stream multiplexing, which shows up on pages that pull many small assets over slow connections.
Two groups need a longer plan. If you build httpd against the standalone mod_h2 from GitHub, versions 2.0.8 through 2.0.42 are affected and no fix is coming. That repository was archived by its owner on June 9, 2026, is read-only, and its README points users at the module maintained inside the Apache httpd project. Migrating to the bundled module is the fix. If you run on a managed hosting panel, the panel's build is the thing that decides: cPanel ships its own ea-apache24 package through EasyApache 4, and CloudLinux has published guidance for its EasyApache users. Ask your provider for the panel package build, and ask whether h2 is on by default in the profile you are assigned.
What a failed attempt leaves behind
Heap corruption that misses shows up as a dead child process. That is the cheapest signal available while no exploit is public, and it is already in your logs.
# Children dying on signals, with timestamps
grep -hE 'child pid [0-9]+ exit signal (Segmentation fault|Bus error)' \
/var/log/httpd/error_log* /var/log/apache2/error.log* 2>/dev/null \
| awk '{print $1, $2, $3}' | uniq -c | tail -20
# Core dumps, if systemd is catching them
coredumpctl list httpd 2>/dev/null | tail -10
# Tell which clients speak HTTP/2 at all: add %H to the access log format
# LogFormat "%h %l %u %t \"%r\" %>s %b %H" h2aware
awk '$NF=="HTTP/2.0"{print $1}' /var/log/httpd/access_log | sort | uniq -c | sort -rn | head
Read the output with two cautions. MaxRequestsPerChild and routine graceful restarts churn children on their own, so a baseline matters more than a count. And a segfault is evidence of a crash, not of an attack; correlate it with a burst of HTTP/2 requests from one source before escalating. Raising mod_http2 to LogLevel http2:info for a day adds stream-level detail if you need to see resets, at the price of log volume.
Check the package, then prove h2 is off
For every Apache host that answers from the internet, answer the three questions today: module loaded, h2 or h2c negotiated on the wire, reachable. Patch through the distribution package and confirm it with the changelog rather than the version string. Where the window is further out, drop h2 and h2c and prove the host stopped negotiating HTTP/2 with an ALPN check from outside. Rebuild container images rather than restarting them. Put any host built against the archived standalone mod_h2 on a migration list this week, because that one has no patch at the end of it. If you want a second set of eyes on the inventory, or on which of your public hosts actually terminate HTTP/2 themselves, that is a short conversation.
Need expert cybersecurity guidance?
We provide hands-on cybersecurity consulting for organizations that need practitioner-level expertise. If you want a second set of eyes on which of your public hosts terminate HTTP/2 themselves, book a session with our team.
