PHP's November fixes close gaps in the PHP2Shell patch; check FastCGI connection reuse first

Two of PHP's 17 advisories from Nov. 19 describe gaps in the October PHP2Shell fix, and a third bug was made reachable by the release that shipped it. Whether your web server reuses FastCGI connections decides which hosts to upgrade first.

nginx.conf with fastcgi_keep_conn on highlighted, beside the order: check reuse, turn off, restart PHP-FPM, upgrade

If you patched PHP2Shell in October, upgrade PHP again: two advisories published Nov. 19 describe gaps in that fix, a third bug was made reachable by the release that shipped it, and the worst bug depends on whether your web server reuses FastCGI connections.

The PHP project released 8.5.13, 8.4.28, 8.3.37 and 8.2.36 on Nov. 19, 2026 as security releases. 8.6.0, a feature release, includes the same fixes. Together they close 17 php-src advisories, 7 rated critical (18 CVEs, counting a bundled uriparser update). PHP said the critical issues "require either non-default configuration or uncommon setups." FPM request smuggling needs no non-default PHP configuration, PHP says; its gate is the front end.

How does this connect to PHP2Shell?

PHP calls the session shutdown bug "an incomplete fix for CVE-2026-109619," the PHP2Shell use-after-free. For the backtrace repopulation bug, PHP says the 8.5.12 fix "did not stop later shutdown work from setting it again." For the magic-method backtrace bug, PHP says 8.5.12 introduced the delayed handling of an allocation failure that makes the bug reachable. The session serializer bug is separate, but PHP shows a gadget-chain RCE on October-patched releases.

Not every critical bug comes with demonstrated RCE. PHP shows code execution for FPM request smuggling, the magic-method backtrace bug, the serializer mismatch (with a gadget chain) and the timezone cache bug (locally), but not for the backtrace repopulation and session shutdown bugs.

Which hosts are exposed first?

Start with the front end. PHP says standard nginx and Apache setups that close the connection after each request are not affected by the demonstrated smuggling attack.

Bug (branches)

You are exposed if

Interim workaround

FPM request smuggling (all supported)

The front end reuses backend connections (nginx fastcgi_keep_conn on, Apache enablereuse=on, HAProxy; LSAPI for disruption only) and an app path opens php://input without reading it, as some frameworks do before routing

Turn reuse off

Magic-method backtrace (8.5 only)

fatal_error_backtraces=1 (8.5 default) and zend.exception_ignore_args=0; php.ini-production sets zend.exception_ignore_args=1

zend.exception_ignore_args=1 or fatal_error_backtraces=0 in the FPM or Apache config, then restart; CLI ini does not count

Backtrace repopulation (8.5 only)

Path-specific overrides plus an early diagnostic from Apache per-directory config or a PHP_VALUE/PHP_ADMIN_VALUE sent by the front end

fatal_error_backtraces=0, restart

Session shutdown (8.3 and later)

A callback hits a fatal error during session cleanup

pm.max_requests=1, at a performance cost

Serializer mismatch; partly overwritten session files (all supported)

Upload progress on with mixed serializers (a runtime change or a shared session store), or an interrupted write to default file sessions; both need a gadget chain

session.upload_progress.enabled=0 in admin config; ini_set() is too late

Timezone cache (all supported)

A startup diagnostic goes to a file error_log

Fix startup warnings or log to syslog, restart

We read fatal_error_backtraces=0 as the one interim setting that covers both 8.5 bugs.

Hosts on 8.1 and older got no upstream release. BleepingComputer says Ondřej Surý published patched packages back to 5.6. PHP's supported-versions page shows 8.2 security support ending Dec. 31, 2026, and 8.4 active support the same day.

What to do now

  1. Platform team, today: search nginx, Apache and HAProxy configs for connection reuse. Turn it off until the host is upgraded, and keep FPM and LSAPI listeners off untrusted networks. We would then restart PHP-FPM: PHP says a smuggled request can change admin-level PHP settings that stay active for later requests on the same worker, and can forge the client address the app's access controls see.
  2. Platform team, this week: upgrade hosts with connection reuse first, then 8.5 hosts not built from php.ini-production, then the rest. Read the effective FPM or Apache ini. Distro builds can backport fixes without changing php -v, so check the distro advisory.
  3. App team, in staging this week: the fixed releases turn session upload progress off by default and reject runtime changes to session settings, including session.serialize_handler, while it is on. An explicit session.upload_progress.enabled=1 keeps the old behavior and the exposure. The fix does not enforce matching serializers across configurations, and separate FPM pools do not get separate session storage automatically. Workers now exit on a failed request shutdown, so expect more restarts in FPM logs.
  4. Vendor management, this week: ask vendors with embedded PHP for fix dates.

What is not known yet

As of 12:02 UTC on Nov. 20, none of the 18 CVEs was in CISA's KEV catalog (version 2026.11.18), and we found no exploitation reports from php.net or vendors. Whether any appliance on the NCSC-NL PHP2Shell list is affected by the FPM bug is unknown.

More in News

All in this topic