Two Critical PHP Flaws Put Millions of Websites at Risk — AI Helped Find Them

On Wednesday, October 21, 2026, PHP developers released four security updates to fix two critical vulnerabilities. The flaws could let hackers run code on vulnerable web servers without a password. AI models from OpenAI and Google played a key role in discovering and investigating the issues.

The vulnerabilities, dubbed PHP2Shell, each carry a severity score of 9.2 out of 10. The PHP Foundation is urging website administrators to update their software immediately.

The news could affect a large part of the internet. PHP underpins popular platforms such as WordPress, Drupal and Joomla. Online stores, government websites and business applications also run on PHP.

Both vulnerabilities are in PHP itself. That means they do not depend on a specific WordPress plugin or a bug in a website’s code.

The first flaw, CVE-2026-109618, occurs when PHP reaches its memory limit under certain conditions. The software can then mishandle internal data, allowing hackers to run their own code on a server.

The second flaw, CVE-2026-109619, also involves memory. After a serious error, PHP may retain pointers to memory that has already been freed. When the same server handles another request, this can open the door to a new attack.

Security researchers have shown that both flaws can lead to remote code execution under certain conditions. This means an attacker can remotely run commands on a vulnerable computer.

On servers running Apache with mod_php, an attack exploiting the first flaw could even succeed before the website’s own software starts running.

A successful attack could give hackers access to files, sensitive data and server functions. The potential damage depends, among other things, on the permissions and configuration of the affected system.

The use of artificial intelligence makes PHP2Shell stand out. Researchers used powerful AI models to analyze complex flaws that had remained difficult to understand for years.

The first signs of one of the problems emerged as early as 2024, when a developer spotted a potential issue with PHP’s memory handling during security work.

At the time, the full severity was unclear. According to the PHP Foundation, a fix was delayed by technical trade-offs, slow decision-making and a shortage of specialized developers. The organization acknowledges that this was a systemic failing.

That changed in August 2026.

On August 11, security researcher Ryan A of Calif found a way to exploit the vulnerability. According to OpenAI, Calif used the GPT-5.6 Sol AI model. OpenAI independently reported the same underlying flaw.

Security firm Wiz also investigated PHP. Its team used Wiz Atlas, GPT-5.6 Sol and Google’s Gemini 4 Argon, among other tools, to find vulnerabilities, test attacks and verify fixes.

According to Google Cloud, Gemini 4 Argon played a key role in uncovering the flaws, developing test attacks and checking the security patches.

The PHP Foundation also used Anthropic’s Claude Mythos for broader security research and OpenAI’s Daybreak Red to review reports and attack methods.

AI did not replace the developers. People investigated the findings, made the technical decisions and worked on the final fixes.

PHP developers have now released fixes for all four officially supported versions.

Google says PHP versions from 5.4 through 8.5.11 are vulnerable to at least one of the flaws.

PHP 8.1 and earlier no longer receive regular security updates from the PHP project. Some Linux distributions and commercial vendors do provide custom security patches. PHP developers have also released limited source-code patches for PHP 7.4, 8.0 and 8.1.

Website administrators can find available versions and technical details in the official PHP security advisory.

For WordPress users, one point is especially important: updating WordPress and its plugins alone is not enough. The vulnerabilities lie in the underlying PHP software running on the server.

Website owners should check which PHP version their hosting provider is running. Servers, containers and other applications with PHP built in also need attention.

The Netherlands’ National Cyber Security Centre (NCSC) has published a dedicated PHP2Shell overview with information about the vulnerabilities and available fixes.

Those unable to update immediately can take additional security measures in the meantime. PHP warns, however, that these do not offer complete protection. Installing a security update remains essential.

The PHP Foundation had originally planned to disclose the vulnerabilities on October 22. After information began circulating on October 20, the project brought the announcement forward by a day.

A video from security firm Calif briefly surfaced, showing a working proof-of-concept attack. Copies then spread across social media.

The PHP Foundation said it had found no working attacks in the wild at the time of the official disclosure. Security firms later reported suspicious activity and attempts to exploit the vulnerabilities. That does not necessarily mean attackers actually took control of any servers.

The threat was serious enough that several organizations took their websites offline temporarily as a precaution.

Above all, the incident shows how AI is changing the cybersecurity landscape.

A complex memory flaw that had seemed difficult to exploit for years was turned into a working attack with the help of modern AI models.

This gives security researchers new ways to uncover dangerous flaws earlier. But attackers can use similar technology to find new ways in faster, too.

The development echoes earlier concerns about AI models capable of carrying out cyberattacks autonomously. Anthropic has also previously restricted access to Claude’s advanced cybersecurity features.

PHP2Shell shows just how important this technology has become. AI can help developers make the internet safer, but it may also leave organizations with less time to protect their systems.