PHP2Shell: AI models helped find critical PHP flaws behind 70% of websites

PHP released patches today, October 21, 2026, for PHP2Shell, a set of critical memory-safety vulnerabilities that Wiz Research says affect software powering an estimated 70% of all websites. OpenAI says its models helped researchers find the flaws. Wiz says it used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol "throughout this research."

That is three tools, at least two labs, one very large attack surface, and different accounts of who deserves the credit on disclosure day. Let us read it properly.

What PHP2Shell actually is

According to Wiz, all of the disclosed bugs are memory-safety issues that stem from how PHP handles fatal errors, particularly out-of-memory errors, while processing requests. When a request ends unexpectedly, the bugs can leave corrupted internal state behind. Later requests or workers can then reuse that state.

The flaws were grouped under two CVEs, with severity reaching CVSS 9.2:

• CVE-2026-109618: an upload-progress unserializer use-after-free plus remote process-memory disclosure. Wiz says it demonstrated pre-authentication remote code execution on WordPress, Drupal and Joomla, along with unauthenticated disclosure of PHP worker memory.
• CVE-2026-109619: a fatal-error backtrace use-after-free plus OPcache type confusion. Wiz demonstrated pre-authentication RCE on Drupal and WordPress.

The detail I would underline is this one: Wiz says exploitation does not depend on a flaw in the application's code, and that "even a simple “Hello World” application running on an affected PHP version can be vulnerable." The usual comfort, that your plugin was fine and the bug was somebody else's, does not apply here. The bug lives in the engine.

PHP powers WordPress, Joomla, Drupal and Wikipedia's MediaWiki, per Wiz. Every PHP version from the last decade, going back to 5.4, is affected by at least one of the vulnerabilities. Wiz also says the issues stem from architecture decisions, "some made over a decade ago." Old code, new eyes.

The timeline

Wiz's disclosure timeline is unusually tidy, so here it is:

• August 19: Wiz Research discovers the first PHP vulnerability using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol.
• August 20: unauthenticated RCE confirmed against Drupal.
• August 21: unauthenticated RCE confirmed against WordPress.
• August 24: Wiz reports the vulnerabilities to the PHP security team.
• September 21: PHP acknowledges the reports and begins developing a coordinated fix.
• October 21: patched releases and public disclosure.

The first bug was found on August 19, and working exploits against two major PHP applications followed within two days. The wait for an acknowledgment from the PHP team took four weeks, and the wait for the patch another four. I am not criticizing the PHP team, since the timeline gives no reasons for the pacing, but it is a useful reminder of which side of the process is quick.

Who gets the credit

The announcements do not agree on who did what.

OpenAI says researchers "operating as part of our Patch the Planet initiative" used OpenAI models to discover the flaws, and that it "led a coalition of maintainers, researchers, and security providers" over two months to prepare fixes and protections before disclosure. The Wall Street Journal, as relayed by Techmeme, reported that PHP2Shell was discovered using GPT-5.6 Sol in August.

Wiz tells a broader story. It credits Gemini 4 Argon with "a particularly significant contribution to finding and helping to remediate the vulnerability." Raluca Ada Popa, Head of Gemini Security at Google DeepMind, said: "We're pleased to see Gemini 4 Argon help defenders bridge that gap, finding vulnerabilities and helping patch them." Wiz thanks the PHP team, Calif and OpenAI for discovery, coordination and patch verification, and presents a multi-party effort. Partners named in the coordination include Automattic, Akamai, AWS, Cloudflare, Fastly, Google Cloud, Imperva, Microsoft and Vercel.

So: OpenAI says it led. Wiz describes collaboration. The Journal headlines GPT-5.6 Sol. Google DeepMind is pleased about Gemini. Each statement may be true, and none of them is an eval.

What was not measured

The cited accounts do not quantify how much each model contributed. Wiz used three tools together, and the material available gives no breakdown between Gemini 4 Argon, GPT-5.6 Sol and Wiz Atlas for discovery versus exploit creation. It also gives no count of how many attempts failed, how much human triage was involved, or how the models compare with conventional fuzzing and review on the same codebase. "AI helped" is a claim about a workflow, not a measurement of a model.

Two other gaps matter more to a site operator. The materials do not say whether any of the vulnerabilities were exploited in the wild before disclosure. They also do not list the specific PHP version numbers that deliver the fix.

What to watch

First, the version numbers: your host's changelog will tell you what you are running faster than any headline. Second, whether exploitation shows up now that the details are public. Third, whether the labs publish anything resembling methodology, because a story about models helping find critical bugs in core PHP deserves more than competing press statements.

The bugs are real, the patches are out, and the 70% figure is an estimate from Wiz and OpenAI. Update first. Argue about the footnotes after.