Dev.to Security 🔐 Cybersecurity 👁 0 📖 7 min read

PHP's November Security Release: 18 CVEs, One Request, and the FastCGI Connection That Never Closed

On November 19, the PHP project published its monthly security release, and it was a heavy one: 18 CVEs fixed across every supported branch, shipped as PHP 8.5.13, 8.4.28, 8.3.37 and 8.2.36. PHP 8.6.0, the new major rele

On November 19, the PHP project published its monthly security release, and it was a heavy one: 18 CVEs fixed across every supported branch, shipped as PHP 8.5.13, 8.4.28, 8.3.37 and 8.2.36. PHP 8.6.0, the new major release, launched the same day with the fixes included. The release arrives barely a month after PHP2Shell, the pair of critical remote code execution flaws that sent the industry into a patching rush compared, at the time, to Log4Shell.

November's batch is different in shape. PHP's security team states plainly that the issues rated critical "require either non-default configuration or uncommon setups." That sentence is doing real work in this release, and understanding it starts with the one bug that comes closest to breaking it.

CVE-2026-129289: smuggling a request through an unread body

The headline flaw is a request smuggling bug in PHP-FPM, the FastCGI process manager that sits behind most nginx and many Apache deployments. It carries a CVSS score of 9.2. According to Upwind's analysis of the release, it gives an unauthenticated attacker PHP code execution with a single request - no memory corruption, no gadget chain, no race condition - which is why Upwind names it the one to close first.

Two preconditions must hold. First, the web server in front must reuse its backend connections to PHP-FPM: nginx with fastcgi_keep_conn on, Apache mod_proxy_fcgi with enablereuse=on, HAProxy FastCGI configurations that retain backend connections, or LSAPI. Second, some reachable application path must open php://input without reading it to the end. Standard nginx and Apache configurations, which close the FastCGI connection after every request, are not affected by the demonstrated attack.

How the bug works, from the fix itself

The mechanics are visible in the fix commit by PHP maintainer Jakub Zelenka, which ships with the release and includes its own regression tests. When a PHP script opens php://input, the engine creates a request body stream without consuming the input the SAPI layer is still holding. At request shutdown, the engine treated the mere existence of that stream as proof the body had been read, and skipped the step where it drains leftover input. The fix changes the condition: the drain now happens whenever the body was not fully consumed, whether or not a stream was ever opened.

The consequence before the fix: when the front end keeps the FastCGI connection open - the FCGI_KEEP_CONN flag - those unread bytes stay in the connection. FastCGI is a framed protocol, so leftover STDIN bytes are parsed as the start of another request. An attacker can embed a complete second FastCGI request inside the body of an ordinary first one, aimed at a script the front-end server would never route to directly, with attacker-chosen parameters. The commit's test files demonstrate exactly this: a hidden request for a smuggled script with a spoofed remote address, riding inside what looks like a normal multipart upload.

On the FastCGI side, the fix adds a second guard: if the previous request's STDIN record was only partially consumed, the next request on that connection is rejected instead of parsed.

The rest of the critical seven

Six more issues carry the release's top rating:

  • CVE-2026-115560: a session serializer mismatch that can lead to RCE. PHP published an informational advisory on October 23, with the full advisory promised for this release. Exposure requires a session.serialize_handler of php_binary or php_serialize combined with a runtime switch to php via ini_set(), plus session upload progress enabled - which is PHP's default. The October PHP2Shell releases did not include a fix for it.
  • CVE-2026-126714: partially overwritten session files can lead to RCE.
  • CVE-2026-129481: a use-after-free after interrupted session shutdown. Per Upwind's analysis, the session cleanup behavior it exploits was introduced in PHP 8.3, so the 8.2 branch is not in scope.
  • CVE-2026-129182 and CVE-2026-129474: two use-after-free flaws tied to PHP 8.5's fatal error backtrace engine - one via repopulation of the last-fatal-error backtrace global after shutdown, one in magic method backtraces. Both apply only to the 8.5 branch. PHP ties 129182, like 129481, to cases the October fix for CVE-2026-109619 did not cover, while 129474 became remotely reachable through a change introduced in 8.5.12.
  • CVE-2026-129496: a timezone cache use-after-free that can lead to RCE.

The other eleven

The remaining fixes spread across the runtime: CVE-2026-129291, an epoll use-after-free in FPM that can lead to local privilege escalation; CVE-2026-129292, an XSS in the FPM status endpoint through unescaped user and script fields; CVE-2026-129538, an out-of-bounds read and write in php-cgi during PATH_INFO correction; CVE-2026-128090, a PHAR OpenSSL signature bypass via algorithm downgrade; CVE-2026-129501, a double free during DOM XML serialization; CVE-2026-15353 and CVE-2026-129506, information disclosure and a use-after-free in SOAP; CVE-2026-128064, mixed-case HTTPS and FTPS URLs that can expose data in plaintext; CVE-2026-127960, a denial of service through folded HTTP response headers; CVE-2026-129438, incorrect RFC 3986 URI serialization for oversized input; and CVE-2026-48841, an upgrade to uriparser 1.0.3.

"Critical" needs context

The PHP security team's caveat - non-default configuration, uncommon setups - is not a reason to delay patching, but it is a reason to patch from understanding rather than panic. Jake Knott, WatchTowr's head of threat intelligence, frames it as three questions every organization should answer first: how is PHP being run (PHP-FPM, Apache mod_php, php-cgi, or something else), how are sessions configured and is session upload progress enabled, and what settings is PHP running with and where are errors logged. In his words, there is "a stark difference in business-as-usual accelerated patching in response to a critical vulnerability, and panic driven by trivially exploitable default configuration vulnerabilities that typically carry the product2shell monikers, which these aren't." WatchTowr expects attackers to probe CVE-2026-115560, CVE-2026-126714 and CVE-2026-129289 first.

The harder problem may be finding PHP at all. Justin Moore, director of adversary operations at Arctic Wolf, points out that few organizations have a complete inventory of where PHP runs: it is baked into CMS platforms, hosting control panels, vendor appliances and old internal applications. His advice is to move internet-facing PHP systems to the front of the patch queue, because attackers have repeatedly shown they move fast on PHP flaws with a path to code execution.

Mitigations while you patch

The advisories ship with workarounds, all described as temporary:

  • For CVE-2026-129289, turn off backend connection reuse on the front end (the settings listed above) or configure it to close the connection after every request. Reading php://input to the end in application code is not a reliable fix on its own: PHP notes that error paths, resource exhaustion and application shutdown can still leave bytes unread. And as a standing rule, do not expose PHP-FPM or LSAPI listeners to untrusted networks.
  • For CVE-2026-115560, set session.upload_progress.enabled=0 - in php.ini, an FPM pool or the server configuration, not via ini_set() or a .user.ini file, because the setting has to apply before request parsing. Then restart every PHP worker. Search the codebase for runtime calls that switch session.serialize_handler to php; applications that never change the handler at runtime can keep upload progress enabled. If this mitigation was already applied in October for PHP2Shell, keep it in place.

None of this replaces upgrading to the fixed release for each branch.

The long tail of old PHP

The release also redraws the support map. PHP 8.2 reaches end of life on December 31, 2026. Everything older is already unsupported, yet still running: Ondrej Sury's distribution packages carried these fixes back to PHP 5.6 on release day, and Aikido published hardened container images with backported fixes for the 8.1/8.0, 7.4, 7.0-7.3 and 5.6/5.5 lines - the image tags keep their original version numbers while containing the November fixes. Upwind adds a reminder that applies well beyond containers: appliances and vendor products that bundle their own PHP build need a rebuild or a vendor update, and they will not show up in a server-level patch run.

The takeaway

Eighteen CVEs in one release sounds like an emergency, and parts of it are: an unauthenticated, single-request code execution path is as serious as web vulnerabilities get. But the distance between "critical on the advisory" and "exploitable in your environment" is configuration - which FastCGI mode you run, how sessions are handled, whether a front end reuses connections. The organizations that can answer Knott's three questions will patch calmly and in the right order. The ones that cannot will learn, again, that the hardest part of a PHP security release is not the patch. It is knowing where PHP is.

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.