
PHP to Shell, Part 5: The CVEs That Carry This Chain in the Wild
Parts 2 through 4 covered the mechanisms. This post is the public record of where those mechanisms actually land: named CVEs, with advisories, exploitation telemetry, and — where I checked it myself rather than taking the original desk research at face value — a note saying so. Every figure below is attributed to a named source and a date, because every one of these numbers moves; EPSS updates daily, attempt counts climb, and a KEV listing from six years ago can still be the most relevant fact about a bug nobody patched.
php2shell: the pair that doesn’t need an app bug at all
Everything else in this post needs something specific to go wrong in an application — an
upload filter, an include() call, an unserialize() sink. CVE-2026-109618 and
CVE-2026-109619 don’t. They’re bugs in PHP’s own runtime, and the practical effect is
unauthenticated RCE against a stock install with no application vulnerability at all.
- CVE-2026-109618: an out-of-memory handling bug. A multipart request that avoids a usable
declared content length can drive the request into an out-of-memory condition while request
memory is only partially initialized; PHP’s OOM path uses
longjmp()to abandon the request without accounting for the early exit, and shutdown then reads or branches on invalid memory. - CVE-2026-109619: a use-after-free across requests. Worker-retained global state can point into memory from the previous request that should have been reset before release; some error or session handling reads the stale pointer before normal initialization runs.
Both were patched out-of-band on October 21, 2026 (PHP 8.2.35, 8.3.36, 8.4.27, 8.5.12), rated
9.2 on CVSS v4. OX Security
and BleepingComputer
both reported exploitation confirmed in the wild, with the highest-risk configuration being
mod_php, where a working trigger fires before any application code runs at all; PHP-FPM,
php-cgi, FrankenPHP and LiteSpeed are also named as affected configurations.
The practical read, stated once and worth repeating for every other row in this post: these two rank above bugs with far higher public name recognition (Drupalgeddon2, a six-year-old ransomware-linked PHP-CGI bug) precisely because they need zero application-level precondition — every other row needs the operator of a specific app to have shipped a specific mistake. php2shell just needs PHP itself.
Tier 1 — the pattern behind almost everything getting mass-exploited right now
Five separate 2026 bugs, one shared shape: an endpoint an anonymous visitor can reach that ends in a file the web server will execute, on a path the framework’s own protections never run across.
- CVE-2026-0740 (Ninja Forms – File Uploads, WordPress, ≤3.3.26, CVSS 9.8): the AJAX upload controller validates nothing about file type before writing the upload, so it bypasses core WordPress’s own extension allowlist entirely — that allowlist never runs, because the write happens through the plugin’s own path. Wordfence’s advisory notes the first patch (3.3.25) was itself bypassable; 3.3.27 is the complete fix, and a partial patch that still gets bypassed is itself a signal attackers keep trying.
- CVE-2026-32475 (Elementor Pro Forms module, ≤4.2.1, fixed 4.2.2): a public form with a File Upload field is all that’s needed — no admin panel, no credentials. Wordfence reported over 190,000 blocked exploitation attempts, with attacks beginning the same day the patch shipped. That’s the highest confirmed attack volume of any row in this post, against a highest-install-base plugin whose paid tier can sit stale behind a license-gated update path.
- CVE-2020-25213 (WordPress File Manager / bundled elFinder connector, <6.9, CVSS 10.0):
the plugin ships an example elFinder connector and renames it to
.phpat install, leaving a live file-write endpoint that needs no authentication at all. Six years on, Shadowserver’s honeypots were still recording hits against this exact CVE in November 2026 — the long tail on a WordPress plugin bug essentially never clears to zero. - CVE-2026-85097 (Bricksforge, ≤3.1.8.9): a two-step TOCTOU — upload a polyglot that passes
validation, then submit a form whose own upload reference gets pointed at a
.phpdestination the validator never saw. - CVE-2026-3844 (Breeze Cache, Gravatar-fetch upload path): the odd one in this list because the vulnerable feature (“Host Files Locally – Gravatars”) is non-default, which should have limited exposure — and still didn’t. It became the primary tool of a webshell-access brokerage operation that reportedly fired it at tens of thousands of targets.
The root cause repeating across all five, independently, across five unrelated vendors: the extension check and the file write are implemented as two separate steps, and only one of them actually runs on the request path an attacker uses. That split — not “unrestricted file upload” as a generic label — is the thing worth looking for in a code review.
Older, still live, worth not dismissing as “patched years ago”
- CVE-2019-11043 (PHP-FPM + nginx, specific
fastcgi_split_path_inforegex): a craftedPATH_INFOunderflows a buffer infpm_main.c, letting an attacker shift which file FPM executes. CVSS 8.7 — the lowest score of anything in the runtime category — but it’s in CISA’s KEV catalog with recorded ransomware use, and it was mass-exploited within days of disclosure using thephuip-fpizdamtool. Low CVSS, high real-world rank: the gap between the two is the whole reason EPSS and KEV exist as separate signals from CVSS. - CVE-2024-4577 (PHP-CGI on Windows, Apache, “Best-Fit” code page conversion): Windows
silently remaps certain characters during request handling, letting an attacker smuggle
PHP CLI flags (
allow_url_include=1,auto_prepend_file=php://input) through a request. KEV-listed with confirmed ransomware use — TellYouThePass began exploiting it two days after the patch shipped, and Censys counted roughly 1,800 ransomware-infected hosts within a week, mostly in China. - CVE-2023-6553 (Backup Migration plugin, filter-chain LFI, unauthenticated, CVSS 9.8): the
same
php://filterchain technique from part 3, but weaponized against a plugin whose ownbackup-heart.phpis reachable directly, outside WordPress’s bootstrap. No upload and no log write needed — it works on read-only, hardened filesystems where most upload-class bugs don’t, which is exactly why it’s worth ranking above several higher-CVSS rows. - CVE-2018-7600 / CVE-2018-7602 (“Drupalgeddon2” and its bypass): Drupal’s Form API accepts
render arrays from request parameters, and a parameter name containing
#survives into the render structure, letting an attacker inject a#post_render/#lazy_buildercallback with attacker-chosen arguments. Mass-exploited within hours of the 2018 disclosure, both KEV-listed with recorded ransomware use, and the archetype for “patch released, mass exploitation same day.” - CVE-2024-34102 (“CosmicSting”, Adobe Commerce/Magento): an unauthenticated XXE in a REST
endpoint reads
app/etc/env.php, including the crypt key, which an attacker then uses to forge admin session material. Sansec documented roughly 4,275 stores breached by seven competing groups, with card-skimmer monetization — and critically, patching alone didn’t evict attackers, because the stolen crypt key stays valid until it’s explicitly rotated.
What actually predicts exploitation, since CVSS mostly doesn’t
Ranked by how well each one actually predicted real-world attacker interest across this set, not by how official it sounds:
- Reachable before the framework’s own protections run. Every upload bug above that got mass-exploited is an endpoint that bypasses the application’s extension allowlist entirely — a separate AJAX controller, a bundled connector, a two-step validate-then-move split. This single property predicts exploitation better than the CWE or the CVSS score.
- A remotely fingerprintable precondition gets exploited; an unfingerprintable one
doesn’t, even at identical severity. CVE-2024-56145 (Craft CMS) needs
register_argc_argv=On— one request confirms it, and it entered KEV within weeks. A WordPress core RCE that needs Imagick specifically backed by Ghostscript (not the far more common GD stack) has zero reported exploitation at the same 8.8 CVSS, because nothing about it is checkable from outside. - Install base, which no scoring system includes. Two identical “unauthenticated upload, CVSS 9.8” plugin bugs can differ by two orders of magnitude in EPSS purely because one plugin runs on far more sites. EPSS absorbs some of that; CVSS absorbs none of it.
- A partial patch is an accelerant, not a mitigation. Ninja Forms’ first patch got bypassed on the way to the real fix, and that bypass is itself a signal that attackers were actively probing the fixed version — not evidence the bug is closed.
- Secret-key compromise outlives the patch. CosmicSting, Laravel’s
APP_KEYpattern from part 4, and Craft CMS’ssecurityKeyall convert a single leaked secret into repeatable access that a patch alone does not revoke. Any remediation runbook for these three needs key rotation as a required step, not an optional one.
Reading this a month from now
Every attempt count, EPSS score, and KEV date above is a snapshot. EPSS specifically moves daily and the relative ranking inside this post can shift without any of the underlying bugs changing. Before using any of this operationally — a patch priority call, a report to a client — pull the current EPSS figure and check CISA’s KEV catalog directly rather than trusting a number written down on November 14, 2026.
Part 6 closes the series where access actually becomes impact: what a webshell looks like once one of these chains succeeds, and which of the detection ideas scattered through this series actually catch it.
Related posts
03 ENTRIEShttps://shulkwisec.github.io/php-to-shell/05-the-cves-that-carry-this-chain/
Printed 2026-11-15


