PHP 8.5.13, 8.4.28, 8.3.37 and 8.2.36 came out on 19 November 2026. Together they fix 18 CVEs, seven of which are rated Critical at CVSS 9.2. The pre-announcement a day earlier said the critical issues “require either non-default configuration or uncommon setups”.
That holds for most of them. It holds less well for two, which are reachable on setups that are ordinary on WordPress hosting, and a third was switched on by default in every PHP build until this release.
None of these are WordPress bugs. They sit underneath it. This is a host-level update, and patching WordPress does nothing for any of it.
What it means for WordPress
The middle column is how likely the bug is to be reachable at all, not how bad it would be if it were. Nothing here is common, and most of it stays rare even with an unusual plugin or code layout.
| Issue | Reachable? | Why |
| Timezone cache use-after-free (CVE-2026-129496) | Uncommon | Decided by server config alone. Nothing in core triggers it and nothing in core prevents it. |
| PHP-FPM request smuggling (CVE-2026-129289) | Uncommon | Every REST request meets the application-side condition, so it comes down to whether FastCGI keep-alive is on. |
| HTTP stream wrapper bugs (CVE-2026-128064, CVE-2026-127960) | Rare | Core never uses the wrapper. Only plugins that skip wp_remote_get(). |
| Session deserialization (CVE-2026-126714, CVE-2026-115560) | Very rare | PHP writes the session files without core’s involvement, but a plugin would have to read them back, and few call session_start(). |
| Magic method backtrace use-after-free (CVE-2026-129474) | Very rare | PHP 8.5 only, the production INI blocks it, and core has 14 magic method definitions in the whole tree. |
| PHP-FPM epoll privilege escalation (CVE-2026-129291) | Post-compromise only | Matters after a plugin bug has already handed someone the worker. |
| SOAP, PHAR, DOM and CGI fixes | Not applicable | Nothing in core reaches them. Plugin territory. |
The short version: two of these deserve attention this week, and the rest are narrow enough to ride along with your normal patching.
All 18 fixes
| Vulnerability | Severity | Branches affected | Requirements |
| Timezone cache use-after-free (CVE-2026-129496) | Critical 9.2 | 8.2, 8.3, 8.4, 8.5 | A startup diagnostic written to a file-based error_log |
| Partially overwritten session files (CVE-2026-126714) | Critical 9.2 | 8.2, 8.3, 8.4, 8.5 | files session handler, attacker-influenced session content, a gadget chain |
| Session serializer mismatch (CVE-2026-115560) | Critical 9.2 | 8.2, 8.3, 8.4, 8.5 | Upload progress enabled, two readers with different session.serialize_handler |
| PHP-FPM request smuggling (CVE-2026-129289) | Critical 9.2 | 8.2, 8.3, 8.4, 8.5 | FastCGI connection reuse on the front end, app opens php://input |
| Use-after-free after interrupted session shutdown (CVE-2026-129481) | Critical 9.2 | 8.3, 8.4, 8.5 | A fatal error inside a session cleanup callback |
| Use-after-free in magic method backtraces (CVE-2026-129474) | Critical 9.2 | 8.5 only | zend.exception_ignore_args=0 and fatal_error_backtraces=1 |
Use-after-free via last_fatal_error_backtrace (CVE-2026-129182) | Critical 9.2 | 8.5 only | Administrative INI config that fails to restore after a request |
| PHP-FPM epoll use-after-free (CVE-2026-129291) | High 8.8 | 8.2.7+, 8.3, 8.4, 8.5 | Code execution already inside a worker, catch_workers_output=yes |
| DoS through folded HTTP response headers (CVE-2026-127960) | High 8.7 | 8.2.28+, 8.3.18+, 8.4.5+, 8.5 | App fetches an attacker-controlled response through the HTTP stream wrapper |
| Mixed-case HTTPS and FTPS URLs sent in plaintext (CVE-2026-128064) | High 8.2 | 8.2, 8.3, 8.4, 8.5 | An HTTPS://-style URL reaching a stream wrapper, plus a network observer |
| PHAR OpenSSL signature bypass (CVE-2026-128090) | High 7.7 | 8.2, 8.3, 8.4, 8.5 | An OpenSSL-signed Phar the attacker can rewrite |
| PHP-CGI out-of-bounds read and write in PATH_INFO (CVE-2026-129538) | Moderate 6.9 | 8.2, 8.3, 8.4, 8.5 | php-cgi only, with cgi.fix_pathinfo=1 and a mismatched mapping |
| SOAP information disclosure (CVE-2026-15353) | Moderate 6.3 | 8.2, 8.3, 8.4, 8.5 | ext-soap parsing untrusted SOAP, and the decoded array returned |
| SOAP use-after-free on DTD entity references (CVE-2026-129506) | Moderate 6.3 | 8.2, 8.3, 8.4, 8.5 | ext-soap parsing attacker-controlled XML, WSDL or schema |
| Double free during DOM XML serialization (CVE-2026-129501) | Moderate 6.3 | 8.4, 8.5 | Dom\XMLDocument or Dom\HTMLDocument plus saveXml() on untrusted input |
| Incorrect RFC 3986 URI serialization (CVE-2026-129438) | Moderate 6.3 | 8.5 only | A URI of roughly 2 GiB passed to Uri\Rfc3986\Uri |
| PHP-FPM status endpoint XSS (CVE-2026-129292) | Moderate 5.3 | 8.2, 8.3, 8.4, 8.5 | Status endpoint enabled and rendered as HTML by an operator |
| Bundled uriparser upgraded to 1.0.3 (CVE-2026-48841) | Not rated | 8.5 only | Dependency refresh in ext-uri |
What the code shows
Session files get written for apps that never start a session
CVE-2026-126714 and CVE-2026-115560 are deserialization bugs in PHP’s own session storage, and core never touches ext-session, so the instinct is that it is not in scope. Half of that is right. php_session_rfc1867_callback() in ext/session/session.c runs during multipart parsing, before RINIT, and does not wait for the application to ask for a session:
case MULTIPART_EVENT_FILE_START: {
...
php_rinit_session(0);
PS(id) = zend_string_init(Z_STRVAL(progress->sid), Z_STRLEN(progress->sid), 0);
progress->sid comes from a form field whose name matches session.name, so an unauthenticated visitor POSTing a multipart body with PHPSESSID and PHP_SESSION_UPLOAD_PROGRESS fields makes PHP write a session file under an ID they picked.
That was the default everywhere: session.upload_progress.enabled is "1" in the 8.4.27 INI table and "0" in 8.4.28, unchanged since the feature shipped in 5.4.0. But writing the file is only half a deserialization bug. Something has to read it back with the php serializer, and core never does, while plugins that call session_start() are rare. What is worth fixing is a shared session.save_path, because separate FPM pools do not get separate session storage, and that is how one tenant’s planted records reach a neighbour.
The timezone cache bug does not care what you run
CVE-2026-129496 has nothing to do with dates in your application. PHP timestamps file-based error_log entries using the configured timezone, and on non-thread-safe builds a diagnostic during module startup can leave the timezone cache allocated in startup memory. That memory is freed when startup ends while PHP keeps a pointer to it, and FPM and mod_php workers inherit the stale pointer.
The trigger is mundane: an invalid date.timezone or a deprecation notice on a session setting, then an early diagnostic on a fresh worker, such as the max_input_vars warning any visitor can provoke with a large form. The researchers crashed an FPM worker on default settings, then demonstrated code execution under both FPM and mod_php with ASLR on. There is no application-level mitigation, because there is no application involvement.
Request smuggling and php://input
CVE-2026-129289 needs the front end to reuse backend FastCGI connections, via fastcgi_keep_conn on in nginx or enablereuse=on with mod_proxy_fcgi, and the application to open php://input without draining it, because opening the stream makes PHP skip the cleanup that discards unread body bytes.
With a malformed multipart body, the leftovers get parsed as a fresh FastCGI request on the same connection, which the advisory says can run attacker-controlled PHP as the worker user, bypass front-end routing and spoof client addresses. WordPress meets the application side unconditionally, in WP_REST_Server::serve_request():
$request->set_file_params( $_FILES );
$request->set_headers( $this->get_headers( wp_unslash( $_SERVER ) ) );
$request->set_body( self::get_raw_data() );
get_raw_data() is a file_get_contents( 'php://input' ) and runs on every REST request, multipart ones included, as set_file_params() two lines above shows. xmlrpc.php and the bundled IXR server do the same. Reading to the end is not a reliable control either, so until you patch, turn off connection reuse.
Stream wrappers: core is clear, plugins may not be
CVE-2026-128064 is a case sensitivity bug in PHP’s built-in http:// and https:// wrappers. The check that turns TLS on only looks for a lowercase s at a fixed offset, so HTTPS://example.com goes out in plaintext, redirect targets included. CVE-2026-127960 is quadratic behaviour in the same wrappers on folded response headers, enough for two malicious responses to occupy both workers of a two-worker pool.
Both need the application to fetch a URL through the wrapper, and core does not. Its two HTTP transports are cURL and WP_Http_Streams, and despite the name the latter opens a raw socket with stream_socket_client() rather than passing a URL to fopen(), while WP_SimplePie_File routes feeds through wp_safe_remote_request(). The exposure is plugins that use file_get_contents() or fopen() on a remote URL instead of wp_remote_get().
Magic methods on PHP 8.5
CVE-2026-129474 is a use-after-free in the arguments PHP collects when building a fatal-error backtrace during a __call() or __callStatic() dispatch, and the advisory confirms remote code execution on 8.5. Two things keep it off most sites. It needs zend.exception_ignore_args=0, and php.ini-production ships that setting as On while php.ini-development sets it Off. Core also defines __call or __callStatic in only 14 places tree-wide. Setting display_errors=0 does not help, because the backtrace still gets built.
The rest
The PHP-FPM epoll use-after-free, CVE-2026-129291, is post-exploitation, which is why it matters. It turns code execution inside a worker, the routine outcome of a vulnerable plugin, into code execution as the FPM master, usually root. It needs catch_workers_output, which is not the PHP default but is enabled in the official PHP-FPM Docker images.
The rest lands outside core. The DOM double free only affects Dom\XMLDocument and Dom\HTMLDocument, the API added in 8.4, while core uses the legacy DOMDocument class and has one saveXML() call, in wp-admin/includes/misc.php, against a config file an administrator owns. Core has no SoapClient reference and loads no signed Phars at runtime, which puts both SOAP bugs in plugin territory and the PHAR signature bypass in CI tooling. The last three need conditions core does not create: the php-cgi binary, a status endpoint rendered as HTML, and a URI of roughly 2 GiB.
So is it safe?
Mostly, with two exceptions. The timezone cache use-after-free is unauthenticated, needs nothing from your application, and its preconditions are a startup diagnostic plus a file-based error log, which is a configuration a lot of sites drift into after a PHP version bump.
The FPM request smuggling bug has an application-side precondition that WordPress meets on every REST request, so whether it can be reached on your server comes down entirely to whether your front end reuses FastCGI connections. That is a setting you can go and read right now.
The rest behaves the way the pre-announcement described, and no plugin update changes anyone’s exposure either way. These are still your problem even though they are not WordPress bugs, and two of them sit close enough to ordinary hosting that the PHP update should not wait for your next maintenance window.
Timeline
Update now
Move to 8.5.13, 8.4.28, 8.3.37 or 8.2.36, whichever matches your branch. PHP 8.1 and earlier are end of life and received nothing here, so every CVE above that applies to 8.2 applies to those permanently. If you cannot update immediately, the highest-value interim steps, in order:
- Set
session.upload_progress.enabled=0administratively, in every pool and virtual host that can write to a shared session directory. Anini_set()in application code is too late, because the write happens before your code runs. - Turn off FastCGI connection reuse on the front end, with
fastcgi_keep_conn offin nginx orenablereuse=offformod_proxy_fcgi. - Clear any startup diagnostics PHP is logging, restart the FPM master and all workers, and consider sending errors to syslog rather than to a file.
- Set
catch_workers_output = noon pools that do not need it, especially on container images that enable it by default. - On PHP 8.5, confirm
zend.exception_ignore_args=Onis actually in effect. Check pool overrides and.user.ini, not just the CLI config. - Give each tenant its own
session.save_path. Separate FPM pools do not get separate session storage on their own.
Then audit the plugins that do what core does not. Fetching remote URLs with file_get_contents() instead of wp_remote_get() is the common one. Talking to SOAP endpoints is less common, and calling session_start() is rarer still, but all three are where whatever application-level exposure you have will be sitting.



