
Differential lab and PoC driver for CVE-2026-119618, an unauthenticated HTTP/2 race in Apache httpd mod_http2 assign_headers() that corrupts the heap and crashes workers.
assign_headers() in mod_http2 writes a synthesised Content-Length: 0 into
the wrong table. It belongs in the request-private r->headers_in; it goes
into req->headers, which lives on the stream pool that the HTTP/2 connection
thread still owns. When that table happens to be full the write grows it,
allocating from a pool whose allocator has no mutex, while the connection
thread is allocating from and freeing to the same allocator. Both threads can
be handed the same memory, and the child process's heap is corrupted.
Reaching it takes one unauthenticated HTTP/2 request shape and no credentials.
Affects httpd 2.4.55 through 2.4.70 and standalone mod_h2 2.0.8 through 2.0.42.
Fixed in httpd 2.4.71 on 11 November 2026. icing/mod_h2 is archived at 2.0.42
with no fixed release, so this targets httpd.
./run-lab.sh build # compile httpd 2.4.70 and 2.4.71, build the runner
./run-lab.sh up # start vulnerable, fixed and mitigated servers
./run-lab.sh diff # classify both, no race driven
./run-lab.sh exploit # drive the race against the vulnerable server
./run-lab.sh probe # every request shape, trigger and control
./run-lab.sh down
The hit rate is dominated by --pad, the length of each filler header value:
./run-lab.sh sweep-pad # find the local optimum, once per host
PAD=1536 ./run-lab.sh exploit # then drive the race at it
On the build below that took the rate from 36.6 to 277.3 events per 10,000
requests. The response is quantised, not proportional — 512 and 1536 are
peaks, 768 and 1024 are dead zones that produce almost nothing. Do not inherit
the number; measure it, because a dead value looks like a working exploit
finding nothing. ./run-lab.sh sweep-streams, sweep-fill and sweep-conn
tune the rest. Read sweep-streams rather than sweep-conn for
concurrency: the contending threads belong to a single connection, so
spreading a fixed request budget over more connections lowers the pressure on
each allocator and the rate falls.
| 0 | not affected |
| 1 | affected build identified, or corruption confirmed |
| 2 | unreachable |
| 3 | no request reached the vulnerable code, so not a valid test |
| 4 | mitigated: HTTP/2 is not offered |
| 5 | affected build, but the race did not land in this run |
5 exists because a probabilistic miss is a property of the run, not a clean bill of health for the target.
Measured on the event MPM, aarch64, engine kernel 7.0.14-linuxkit, 2 CPUs per container, 14 November 2026. Every figure below is reproducible with the commands above.
Events are sanitizer reports naming this code path, or children killed by a fatal signal, per 10,000 requests the server's access log shows reached a handler. Five matched runs per arm, 15,000 requests offered to each:
| build | reached handler | events | per 10,000 |
|---|---|---|---|
| vulnerable 2.4.70 | 61,599 | 380 | 61.69, SD 4.14 |
| fixed 2.4.71 | 72,478 | 0 | 0.00 |
A full table turns out to be necessary but not sufficient: the growth must
also be large enough to miss apr_palloc()'s fast path, so 8, 18 and 38
filler headers fill the table and produce nothing while 78 and 158 corrupt.
patches/ holds two diagnostic instruments: an allocator-level concurrency
assertion for APR, and a window-widening delay that must be applied to both
builds or it proves nothing. Both supplement the unmodified-binary result.
| Field | Value |
|---|---|
| CVE | CVE-2026-119618 |
| Component | Apache HTTP Server, mod_http2 (standalone mod_h2) |
| Affected | httpd 2.4.55 - 2.4.70; mod_h2 2.0.8 - 2.0.42 |
| Fixed in | httpd 2.4.71; mod_h2 archived at 2.0.42, no fixed release |
| Fix commit | 824beab9b9424d6cfb29595248fdbf6c3dddeab1 (Eric Covener, PRs #975/#977) |
| CWE | CWE-362 (ASF) |
| Vector | network, unauthenticated, HTTP/2 over h2c or TLS h2 |
| Workaround | remove h2 and h2c from Protocols |
| Published | 11 November 2026 |
| Role | Who |
|---|---|
| Discovered and reported | Anthropic's OSS Scanner (ANT-2026-MAPKT6MZ); Anthony Weems and Simon Scannell, Google Cloud Vulnerability Research |
| Vendor advisory and fix | The Apache HTTP Server project, httpd 2.4.71 |
| Reference PoC and lab design | googlesky (Le Hieu), MIT |
| Reference root-cause analysis | blkph0x, MIT |
| This implementation | d4ytox |
The vulnerability is not mine. The HTTP/2 client, the PoC driver, the lab build, the controls and the sweeps are. The request-shape taxonomy and the functional-probe idea follow googlesky's public lab; the root-cause reading cross-checks against blkph0x's analysis. Both are MIT licensed and credited in the file headers.
824beab9b9424d6cfb29595248fdbf6c3dddeab1 in httpd 2.4.71MIT, see LICENSE. The grant covers this implementation. The vulnerability was found by the researchers named above; the reference PoC and analysis are MIT-licensed work by googlesky and blkph0x, credited in the file headers and the Origin table, and this is not a relicensing of their work.
For authorised testing and research only. The PoC drives a heap-corruption race that crashes the target's httpd child processes, so it is disruptive by nature, and under the event MPM each crashed child takes every connection it was serving with it. Running this against a system you do not own, or do not have written permission to test, is illegal in most places. The lab is self-contained and reachable only from inside the container engine.