AI Turned a Decade-Old PHP Flaw Into an Internet-Scale Crisis

AI Turned a Decade-Old PHP Flaw Into an Internet-Scale Crisis

Two 9.2-Rated Flaws Force an Emergency Release

PHP shipped four emergency security releases on October 21, patching two critical vulnerabilities tracked as CVE-2026-109618 and CVE-2026-109619. Both carry a CVSS v4.0 score of 9.2. The Next Web and SecurityWeek report that an attacker needs no credentials at all to execute code on a PHP server, and the research community has collectively nicknamed the pair PHP2Shell.

The Bugs Live Inside the Engine

The flaws sit in PHP's own memory management and error handling, code that nearly every request touches. CNN and France 24, citing Wiz, report that PHP powers roughly 70 percent of the world's websites, while SecurityWeek notes that affected releases stretch back to version 5.4. Forbes points out that the language sits underneath WordPress, Drupal and Magento, so careful application code buys nothing here.

AI Lowered the Exploitation Barrier

Wiz researchers wrote that the underlying flaws had been present for decades, but demonstrating exploitability required an exceptionally complex chain of steps, at the edge of what human researchers could practically achieve. Forbes quotes OpenAI calling this one of the most impactful security issues in recent memory, and framing it as a Defender's Window. The Wall Street Journal quotes Wiz co-founder Ami Luttwak calling PHP2Shell the most severe AI vulnerability ever found.

Four Fixed Versions, One Choice

The patched releases went live at roughly 14:00 UTC on October 21: versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. SecurityWeek and BleepingComputer both report the PHP Foundation's advisories stating there are no complete mitigations and updating is the only safe choice. QZ and SecurityWeek note that the 8.1 branch and older are end of life and receive no official fix, though some Linux distributions and long-term-support vendors ship backports.

A Leak Forced the Patch Out Early

BleepingComputer reports that watchTowr CEO Benjamin Harris said the two flaws were embargoed until11 a.m. ET on Thursday. The Next Web and Quartz report that details surfaced on Hacker News on October 20 before being removed, pushing the PHP Foundation to move the release up by 24 hours. The Wall Street Journal records Calif CEO Thai Duong warning that the moment patches appear, someone could run AI against them and post an exploit script.

Attackers Can Also Read Credentials

Wiz's technical write-up shows attackers can read PHP process memory without first achieving code execution, pulling passwords, tokens and API keys. SecurityWeek reports that the two CVE identifiers group four memory-safety issues but do not cover every variant, and that both tracked and likely untracked variants exist.

Websites Went Dark on Every Continent

SecurityWeek and France 24 report that Dutch organisations took public-facing sites offline, including the government portal Overheid.nl, DigiD, the Dutch parliament site, the national cybersecurity centre and the data protection authority. The Economic Times adds the Port of Rotterdam, football association KNVB, the five largest political parties and several hospital websites. France 24 and Daily Sabah report that French local government sites, including the Paris police site, were unreachable at the request of the national cybersecurity agency. A Reuters factbox reports the Wikimedia Foundation said all of its PHP systems were affected and it took a number of them down as a precaution.

Agencies Sound the Alarm

France 24 reports the US National Security Agency instructed organisations to patch immediately or pull vulnerable systems offline, saying the flaws often provide a direct path to the server's file system or database. The New Straits Times reports CISA advised organisations to update immediately and had not observed exploitation as of publication. USA Today reports the White House was notified on Friday, October 16. USA Today and France 24 add that cybersecurity agencies across Europe, Australia, Japan and Britain urged users to update or go offline, while The Economic Times notes Australia's agency was not aware of active exploitation.

A Two-Year-Old Warning

Officechai reports that PHP's first external security audit, in 2024, was funded by Germany's Sovereign Tech Agency and carried out by Quarkslab, where a contracted core developer spotted early warning signs of the memory-corruption issue while fixing something else, yet nobody could turn it into a working exploit. CrowdStrike's analysis is more granular: one triggering path came from a performance optimization introduced in October 2017, first shipped in PHP 7.3.0 in December 2018, meaning the exploitable window has existed for more than seven years.

August: AI Closed the Loop

Officechai reports that on August 11, 2026, a researcher at Calif reported an unconditional remote-code-execution path using GPT-5.6 Sol, OpenAI reported the identical issue independently the same day, and Wiz later contributed further attack vectors. Forbes and France 24 report that Wiz discovered the flaws independently using Google's Gemini 4 Argon alongside GPT-5.6 Sol, all as part of OpenAI's Patch the Planet program.

Exploitation Has Already Started

SecurityWeek cites Hadrian, which reported php2shell exploited in the wild at scale, with most observed attacks abusing PHP's session upload-progress path and more exploit paths expected before and after the fix. QZ reports Cloudflare rolled out firewall rules to every customer, including free-tier users, and that CTO Dane Knecht said the company is seeing early attacks.

Upgrade or Go Dark

SecurityWeek and BleepingComputer both warn that updating WordPress or Laravel does not update the PHP runtime underneath, which comes from the server package, the host or the container image, and that operators must then restart Apache, PHP-FPM and any long-running PHP workers. Forbes describes website owners scrambling to patch. The Wall Street Journal documents how researchers kept the flaw closely guarded while they worked on fixes, for fear an AI-aided attacker would find and exploit it first.