Apache recently released version 2.4.71 of its HTTP server, the piece of software that still serves much of the open web, and with it closed a critical flaw in the module that handles HTTP/2. The identifier is CVE-2026-119618 and the stated scope runs from 2.4.55 to 2.4.70, almost three years of releases. For its part, Red Hat scores the problem 9.8 out of 10 and maintains that an attacker can execute commands on the server without presenting credentials, provided HTTP/2 is active.
The severity does not come from the type of flaw, which is a known one, but from the combination of two factors. The first is that HTTP/2 is here to stay: it is enabled with a directive and almost nobody reviews it after installing it. The second is that the error appears when two requests coincide in time, which turns any server with real traffic into a testing ground for anyone who knows how to provoke that coincidence. A server that only speaks HTTP/1.1 is not affected, and that detail is worth more than any headline.
The other half of the story is the calendar. Red Hat, Debian and Ubuntu published packages on 11 and 12 November; the module distributed separately on GitHub still has no fixed version, and that gap decides who is exposed this week.
The flaw is in how mod_http2 allocates memory
The official Apache advisory describes the problem as a concurrent execution vulnerability with incorrect synchronization, classified CWE-362, within the assign_headers() function of the mod_http2 module. In plain terms: when several requests arrive at once, the module writes the response headers into a connection memory space that is allocated through a lock-free allocator. Two streams can write to the same slot at the same time and the process memory stack ends up corrupted.
Stack corruption is not a cosmetic defect. If the attacker controls part of the content being written, they can place the bytes they want in that area and divert the process execution toward their own code. The theoretical result is remote execution: the server stops being a server and starts running whatever it is sent.
Apache rates the flaw as critical and offers no mitigation in the text of the advisory: the recommendation is to update to 2.4.71, and anyone who cannot yet do so has the option of no longer offering HTTP/2.
The cause of the exposure
The HTTP/2 option does not enable itself. The Protocols directive accepts the value h2 for HTTP/2 over TLS and h2c for HTTP/2 over unencrypted TCP, and it can be set globally or within a specific virtual host. An administrator who wants to know whether their machine is exposed must check both locations, because an inherited virtual host can keep the module active even if the general configuration no longer mentions it.
The rest of the chain has its own trap. The module distributed separately, published on GitHub as mod_h2, carries the same error in versions 2.0.8 to 2.0.42. That repository has been archived since 9 June 2026 and its last release is precisely the last vulnerable one, so anyone using it cannot expect a release: they must migrate to the project module or turn off HTTP/2.
| Platform |
Status |
Package or advisory |
| Apache HTTP Server |
Fixed on 11 Nov |
2.4.71 |
| RHEL 9 |
Fixed |
mod_http2-2.0.26-7.el9_9.1 (RHSA-2026:98981) |
| RHEL 10 |
Fixed |
mod_http2-2.0.29-5.el10_3.1 (RHSA-2026:98978) |
| Debian stable and testing |
Fixed in security |
DSA-6620-1 and DLA-4902-1, version 2.4.71 |
| Ubuntu 20.04 to 26.10 |
Fixed |
USN-9090-1 |
| Amazon Linux 2 and 2023 |
Pending |
no package published |
| mod_h2 from GitHub |
No patch |
last release 2.0.42, archived architecture |
The table makes uncomfortable reading for anyone who trusts the distribution. On the day of the announcement, Red Hat already had packages for its two enterprise branches, while Amazon Linux still had no date. Debian's security repositories published the fix with the advisories DSA-6620-1 and DLA-4902-1, and package tracking still flags the base versions of bookworm and trixie as vulnerable, which in practice means that a system updated only through the normal channel can fall behind.
Ubuntu resolved the issue with the security notice USN-9090-1, published on November 12, which covers from 20.04 to 26.10 and leaves 2.4.68-1ubuntu6.2 as the most recent release. Older versions keep the fix behind Ubuntu Pro, a difference that matters if the installation is 18.04 and has no subscription.
Exploitation confirmed, not yet observed
The Red Hat advisory states that there is public research with a reliable exploit against a default server with mod_http2 enabled, although it does not identify that work, does not link to it and provides no details to reproduce it.
The rest of the ecosystem lags behind. The security firm Wiz said on November 11, that there was no exploit available and that no cases of exploitation in progress had been recorded, and its own exposure count serves to size the problem without dramatizing it: 64% of the cloud environments analyzed had resources with affected versions and 51% also had them exposed to the internet, figures that do not discount each server's configuration. The proof of concept published on GitHub reproduces memory corruption and process crashes, not remote code execution.
The detail that separates an advisory from an emergency is another matter. CVE-2026-119618 did not appear in CISA's catalog of exploited vulnerabilities in the November 10 review, and the count of the twelve flaws the agency added in ten days shows the bar it does use to move federal agencies. Nor are there documented campaigns against this flaw, and writing it as if there were would be inventing them.
The same release fixes two other problems
Updating to 2.4.71 does not only close the remote code execution. The package fixes mod_md with CVE-2026-106023, a null pointer dereference that crashes the child process while handling requests, and mod_vhost_alias with CVE-2026-109163, a single-byte write on the stack that Apache classifies as low. Neither justifies a maintenance window on its own, but both travel in the same download.
The mod_vhost_alias case is interesting for another reason: it comes immediately after CVE-2026-63292, an overflow in the same module that was fixed in 2.4.69, released in October. Two flaws in the same component in two consecutive releases point to code with little test coverage, and there the lesson is not the specific patch but the module.
One more data point is worth noting on the question of relative urgency. The 2.4.x branch has accumulated 44 releases, and in 2.4.69 alone 20 vulnerabilities were fixed, according to the project's own changelog, and those two figures sum up the real pace of a component that almost nobody reviews.
That pace comes at a cost. Anyone who updates every six months lives in permanent debt, and this autumn it has paid for it in the form of eighth Chrome zero-day. El the FreeSWITCH and Apache Juneau case confirms the pattern with two serious flaws in the same week.
Removing HTTP/2 costs traffic and will cost much more
The mitigation Wiz recommends for now is to remove h2 and h2c from the Protocols directive and leave the server on HTTP/1.1 until the problem is fixed. But this has a measurable cost: HTTP/2 multiplexing disappears, each resource again needs its own connection and performance on high-latency networks suffers. On a server with thousands of daily visits, that price is paid in milliseconds and bandwidth, not in security.
The alternative, updating immediately, depends on a package being available for your system. Here this week's map leaves two gaps: Amazon Linux and mod_h2. For the first, you have to wait or compile; for the second there will be no release, because the project is archived and its last deliverable is precisely the vulnerable one.
It is also worth debunking an idea circulating in administration forums. A WordPress patch does not address this flaw, because the error lives in the web server and not in the application; the Apache advisory does not mention any CMS. The same applies to hosting panels that wrap httpd without updating the underlying binary.
The calendar rules. Every day a server with h2 enabled remains on 2.4.70 or earlier, the distance between a proof of concept that only crashes processes and an exploit that executes code depends on whether someone publishes the second step. That publication has not happened yet, and the outstanding work is to make sure you are not caught on the old version when it does.