On 19 November 2026 the PHP team shipped security releases for every supported branch: PHP 8.5.13, 8.4.28, 8.3.37 and 8.2.36. They fix 18 CVEs, and the most serious one, CVE-2026-129289, is a critical PHP-FPM request smuggling flaw that can let an unauthenticated attacker run PHP code on your server with a single request, if your setup matches a specific pattern.
This post focuses on that pattern. You’ll learn how the bug works in plain English, how to check whether your nginx, Apache or HAProxy config is exposed, how to mitigate it today, and what else is in the release. Sources: the official PHP advisory GHSA-qc49-hpvq-g2w6, Aikido’s write-up and BleepingComputer’s coverage.
Fixed versions at a glance
| Branch | Fixed in |
|---|---|
| 8.5 | 8.5.13 |
| 8.4 | 8.4.28 |
| 8.3 | 8.3.37 |
| 8.2 | 8.2.36 |
PHP 8.1 and older get no upstream fix. If you are still on them, treat them as affected and plan an upgrade.
How CVE-2026-129289 works
Your web server talks to PHP-FPM over FastCGI. FastCGI connections can carry more than one request. Before PHP accepts the next request on the same connection, it must either read and discard the rest of the previous request body or close the connection.
According to the advisory, PHP reads php://input lazily. If an application opens php://input but does not read it to the end, PHP wrongly treated the open stream as proof the body was handled and skipped the cleanup that discards unread bytes. An attacker can send a malformed multipart/form-data body that stops PHP’s parser at a predictable point and hides a complete FastCGI request inside the leftover bytes.
If the front end keeps the backend connection open, PHP-FPM then reads those leftover bytes as a new, forged request that your web server never created or authorised. The forged request can set FastCGI parameters such as SCRIPT_FILENAME, REMOTE_ADDR and PHP_ADMIN_VALUE, which is enough to run attacker-supplied PHP code as the worker user. Changed settings can even stick for later requests on the same worker.
The advisory notes the bug has existed since PHP 5.6.0. It needs no memory corruption and no race condition.
Two conditions must both be true
- Your front end reuses FastCGI connections to PHP-FPM (or LSAPI).
- Some code path opens
php://inputwithout reading it fully. Frameworks that create an input stream before routing or authentication can trigger this even on endpoints that ignore the body.
Standard nginx and Apache configs that close the FastCGI connection after each request are not affected by the demonstrated attack.
Check your configuration
nginx
The risky directive is fastcgi_keep_conn on;. It is off by default.
# search all nginx config for keep-alive to FastCGI
sudo nginx -T 2>/dev/null | grep -n "fastcgi_keep_conn"
If you see fastcgi_keep_conn on;, you are in the exposed group. It is usually paired with an upstream block that has a keepalive directive:
upstream php {
server 127.0.0.1:9000;
keepalive 16; # often used together with fastcgi_keep_conn
}
location ~ \.php$ {
fastcgi_pass php;
fastcgi_keep_conn on; # <-- exposed until PHP is patched
include fastcgi_params;
}
Apache (mod_proxy_fcgi)
Look for enablereuse=on on FastCGI proxy definitions:
grep -RIn "enablereuse" /etc/apache2/ /etc/httpd/ 2>/dev/null
HAProxy and LSAPI
HAProxy FastCGI applications that keep backend connections, and LiteSpeed’s LSAPI, are listed as relevant configurations. For LSAPI the advisory says the bug can desynchronise and disrupt requests, but it did not establish code execution through LSAPI.
Fix it: upgrade PHP
Upgrading is the real fix. Some common ways:
# Debian/Ubuntu with the ondrej/php PPA or Sury repo
sudo apt update && sudo apt install --only-upgrade 'php8.4*'
sudo systemctl restart php8.4-fpm
# Docker: pull the new tag and rebuild your image
docker pull php:8.4-fpm
docker build --pull -t my-app .
Then confirm the running version, not just the installed one:
php -v
php-fpm8.4 -v # binary name varies by distro
Remember that PHP-FPM keeps running the old binary until it is restarted, and long-lived containers keep the old image until they are redeployed.
Mitigate until you can upgrade
For CVE-2026-129289: stop reusing backend connections
Turn connection reuse off on the front end:
location ~ \.php$ {
fastcgi_pass php;
# remove "fastcgi_keep_conn on;" (default is off)
include fastcgi_params;
}
# Apache: remove enablereuse=on or set it off
<Proxy "fcgi://127.0.0.1:9000" enablereuse=off>
</Proxy>
Reload the web server afterwards (sudo nginx -t && sudo systemctl reload nginx). The advisory says that reading php://input to the end in your code prevents the demonstrated condition, but it is not reliable as the only mitigation, because error paths, resource limits and shutdown can still leave bytes unread.
Also make sure PHP-FPM and LSAPI listeners are never reachable from untrusted networks. Bind FPM to a Unix socket or 127.0.0.1:
; /etc/php/8.4/fpm/pool.d/www.conf
listen = /run/php/php8.4-fpm.sock
For CVE-2026-115560: session upload progress
Another critical fix in this batch, CVE-2026-115560, involves a session serializer mismatch. Aikido explains that an app is only exposed if it configures session.serialize_handler as php_binary or php_serialize and then switches it to php with ini_set() before session_start(), while session upload progress is enabled (the default). The suggested stopgap:
; php.ini or FPM pool config, not .user.ini and not ini_set()
session.upload_progress.enabled = 0
Restart all PHP workers after changing it. Then search your code for runtime handler switches:
grep -RIn "session.serialize_handler" app/ src/ config/ 2>/dev/null
What else is fixed
Aikido lists all 18 CVEs. Highlights include:
- CVE-2026-129474 – use-after-free in magic method backtraces (possible RCE).
- CVE-2026-129496 – timezone cache use-after-free (possible RCE).
- CVE-2026-126714 – partially overwritten session files (possible RCE).
- CVE-2026-129291 – PHP-FPM epoll use-after-free (local privilege escalation).
- CVE-2026-129292 – XSS in the FPM status page.
- CVE-2026-128090 – PHAR OpenSSL signature bypass via algorithm downgrade.
- CVE-2026-128064 – mixed-case HTTPS/FTPS URLs could be handled as plaintext.
These releases also include the October PHP2Shell fixes, so upgrading covers both.
Patch checklist
- Find every PHP runtime: servers, containers, CI images, cron hosts.
- Upgrade to 8.5.13, 8.4.28, 8.3.37 or 8.2.36 and restart PHP-FPM.
- Until then, disable
fastcgi_keep_conn/enablereuseand setsession.upload_progress.enabled = 0. - Keep FPM on a Unix socket or localhost only.
- Re-check
php -von the live process after deploys.
Takeaway
Most default setups are not exposed to the headline PHP-FPM bug, but performance-tuned configs with FastCGI keep-alive are, and the impact is remote code execution. Check your web server config today, turn off backend connection reuse if you need time, and roll out the 19 November releases. The PHP advisory has the full technical details.
