Contents
Apache httpd 2.4.71, released on November 11, fixes CVE-2026-119618, a heap corruption bug in mod_http2 that the httpd project rates critical and describes as leading to remote code execution. It affects httpd 2.4.55 through 2.4.70 with HTTP/2 enabled. The fix commit changes one argument in one function call.
The part that matters for anyone checking their fleet is what the fix did not change. The mod_http2 bundled with 2.4.70 and the one bundled with 2.4.71 both define MOD_HTTP2_VERSION "2.0.42", and both log that string at startup. The advisory lists standalone mod_h2 releases 2.0.8 through 2.0.42 as affected. So a patched 2.4.71 server reports a module version that sits inside the affected range, and a vulnerable 2.4.70 server reports the same thing.
The bug is in which table gets the header
When an HTTP/2 request arrives, mod_http2 parses it on the connection thread (httpd’s code calls this connection c1) into an h2_request that lives in the stream’s memory pool. It then hands the request to a worker thread, which runs it on a secondary connection (c2) through the normal httpd request pipeline. On the worker, h2_create_request_rec() builds a request_rec and calls assign_headers(). This is the 2.4.70 code, from modules/http2/h2_request.c:
static void assign_headers(request_rec *r, const h2_request *req,
int no_body, int is_connect)
{
const char *cl;
r->headers_in = apr_table_clone(r->pool, req->headers);
...
if (!is_connect) {
cl = apr_table_get(req->headers, "Content-Length");
if (no_body) {
if (!cl && apr_table_get(req->headers, "Content-Type")) {
/* If we have a content-type, but already seen eos, no more
* data will come. Signal a zero content length explicitly.
*/
apr_table_setn(req->headers, "Content-Length", "0");
}
}
The function copies the stream’s headers into r->headers_in on the request pool, which is the right thing to do. A few lines later it writes Content-Length: 0 into req->headers, the original table that still belongs to the stream. The 2.4.71 fix changes that call to apr_table_setn(r->headers_in, "Content-Length", "0").
Two things were wrong with the old line. It modified memory the worker thread doesn’t own, which is the security bug. It also never did its job, because r->headers_in had already been copied from req->headers before the write. The request being served didn’t see the Content-Length: 0 it was supposed to get.
Why a header write corrupts the heap
APR tables are arrays allocated from a pool. If apr_table_setn() adds a key and the array is full, APR allocates a bigger array from the pool that owns the table. Here that pool is the stream pool.
In 2.4.71’s h2_session.c, the session pool is created with apr_pool_create(&pool, c->pool) and each stream pool with apr_pool_create(&stream_pool, session->pool). Neither call passes its own allocator, so both inherit the main connection’s allocator, and the connection thread uses that allocator while it keeps reading frames for other streams.
mod_http2 knows this is unsafe. The comment above h2_mplx_c1_create() in h2_mplx.c says the multiplexer pool “needs its own allocator, since apr_allocator_t are not re-entrant”, and the per-request transit pools for workers are created with apr_allocator_create() for the same reason: “This is the only way to have the processing independent of its parent pool in the sense that it can work in another thread.” assign_headers() broke that rule. Red Hat’s description of the root cause matches the code: “a worker thread writes into memory owned by the connection thread through a shared, lock-free allocator.”
The code path takes very little to reach:
no_bodyis true when the stream has no input beam, which inc2_process()means the client setEND_STREAMon theHEADERSframe.- The request carries
Content-Typeand noContent-Length. - HTTP/2 is negotiated, via
h2over TLS orh2cover cleartext.
None of that needs authentication or a particular handler, and a client controls how many headers it sends, so it controls whether the write has to grow the array. Getting from there to code execution means winning a race against the connection thread’s own allocations, and I haven’t seen public exploit code. Red Hat says “public research has demonstrated a reliable exploit” against the default configuration, according to The Hacker News, without naming the research. Wiz says no exploit is available and no in-the-wild exploitation is known. Those two claims aren’t independently verified, and they don’t fully agree, so plan as if exploitation is practical.
The advisory credits Anthropic’s OSS Scanner (reported October 26) and Anthony Weems and Simon Scannell of Google Cloud Vulnerability Research (reported November 3).
Version numbers you can’t use
The httpd version is the only reliable signal for upstream builds, and it only works if you build from upstream. Here are the version strings in play:
| What you look at | Vulnerable example | Fixed example |
|---|---|---|
| Upstream httpd | 2.4.70 | 2.4.71 |
MOD_HTTP2_VERSION in upstream httpd, logged as mod_http2 (v...) at LogLevel info | 2.0.42 (in 2.4.70) | 2.0.42 (in 2.4.71) |
| Standalone icing/mod_h2 | 2.0.8 through 2.0.42 | no release |
RHEL 9 mod_http2 package | earlier builds | mod_http2-2.0.26-7.el9_9.1 |
RHEL 10 mod_http2 package | earlier builds | mod_http2-2.0.29-5.el10_3.1 |
Debian bookworm apache2 | 2.4.68-1~deb12u1 | 2.4.71-1~deb12u1 |
Debian trixie apache2 | 2.4.68-1~deb13u1 | 2.4.71-1~deb13u1 |
I checked the MOD_HTTP2_VERSION row against the 2.4.70 and 2.4.71 tags in the httpd repository: modules/http2/h2_version.h defines "2.0.42" in both, and the startup log line in mod_http2.c prints that macro. The 2.4.x branch now carries a change that makes mod_http2 report the httpd version instead (commit a90e3e5), but it isn’t in 2.4.71.
The standalone repository is a separate problem. icing/mod_h2 was archived on June 9 and is read-only, so anyone who built a newer mod_h2 from GitHub against an older httpd has no fixed release to move to. Wiz’s page lists 2.0.42 as patched in its table and says to upgrade past 2.0.42 in its remediation text. The httpd advisory says 2.0.42 is affected, which is correct for the code.
Distribution packages need the vendor’s tracker, not a version comparison. RHEL ships mod_http2 2.0.26 and 2.0.29 with the fix backported, both inside the affected range, and the httpd version on those systems never reads 2.4.71. Debian’s security tracker shows bookworm and trixie fixed in the security suites with 2.4.71 builds, while the point-release versions (2.4.68) remain vulnerable. Amazon Linux was listed as pending a fix when The Hacker News published.
Why RHEL rates it lower
Red Hat rates the CVE Critical in general but Important for RHEL 9 and 10, because its packages “allocate each HTTP/2 stream from a separate allocator,” which it says “substantially increases the complexity of exploiting the flaw.” That matches the analysis above. If the stream pool has its own allocator, the worker’s write no longer races the connection thread’s allocations from other streams, although it is still writing into a table it doesn’t own.
That’s a vendor hardening choice with an exact effect on the vendor’s own builds. It doesn’t apply to upstream builds, container images built from upstream source, or other distributions.
What to check
First, find out whether HTTP/2 is actually on. The module has to be loaded:
apachectl -M | grep http2
and a Protocols directive has to include h2 or h2c, either server-wide or in a <VirtualHost>. Grep the whole config tree, including included files, because one vhost is enough.
From outside, curl prints the negotiated protocol:
curl -sS -o /dev/null -w '%{http_version}\n' --http2 https://your-origin.example/
2 means HTTP/2 was negotiated. Run this against the origin itself. If a CDN or load balancer terminates TLS in front of Apache, you’re seeing the edge’s protocol. Whether Apache is exposed then depends on whether that proxy talks HTTP/2 to the backend, which you have to check in its config.
Then patch: upstream 2.4.71 or your vendor’s fixed package, confirmed through the vendor’s advisory or package changelog. The Apache advisory offers no workaround. If you can’t patch yet, Wiz’s interim mitigation is to remove h2 and h2c so that only Protocols http/1.1 remains, in every vhost. Browsers fall back to HTTP/1.1 through ALPN. Clients that use h2c with prior knowledge will break. Expect more connections per client, since HTTP/1.1 has no multiplexing.
When this is over, look at how your scanner decided which servers were affected. If it read mod_http2 (v2.0.42 from a log, or compared a RHEL package version against 2.0.42, it got this one wrong in both directions.
Sources
- Apache httpd 2.4 vulnerabilities: CVE-2026-119618, CVE-2026-106023, CVE-2026-109163
- httpd commit 824beab: mod_http2: assign_headers update table
- httpd source at the release tags:
h2_request.c(2.4.70),h2_version.h(2.4.70),h2_version.h(2.4.71),h2_session.c(2.4.71),h2_mplx.c(2.4.71) - Red Hat: CVE-2026-119618
- Debian security tracker: CVE-2026-119618
- icing/mod_h2 releases
- Wiz: CVE-2026-119618
- The Hacker News: Critical Apache HTTP/2 Flaw Could Let Unauthenticated Attackers Execute Code Remotely