PHP Patches Two Critical Remote Code Execution Flaws That AI Models Helped Find

The PHP project on October 21, 2026 released PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two critical flaws that can let an unauthenticated attacker run code on servers running PHP. Researchers at Calif, OpenAI and Wiz reported the full impact with help from frontier AI models, and the PHP Foundation told every PHP user to update now.

PHP is the language behind WordPress, Drupal and MediaWiki, and Cloudflare, citing W3Techs, says it runs on almost 70% of websites. With that reach, and AI models at the center of the discovery, the release quickly became an AI security story: European government, hospital and university sites went offline as a precaution (Reuters via The Next Web), and OpenAI called it "one of the most impactful security issues in recent memory."

What exactly did PHP fix?

The October 21, 2026 releases fix CVE-2026-109618 and CVE-2026-109619, two memory bugs in PHP's core that the project rates 9.2 out of 10 and says can allow unauthenticated remote code execution, according to the php.net advisory.

The first is triggered when a request hits PHP's memory limit at the wrong moment, leaving half built data that PHP later treats as valid. The second lets internal state outlive its memory, so a later request can use freed memory. Both sit in code nearly every request touches, the PHP Foundation wrote.

The project says every PHP installation serving web requests is at risk, most of all Apache with mod_php, where a request can trigger the attack before any application code runs. Wiz, which named the exploit chain PHP2Shell, says versions back to PHP 5.4 are affected and that it showed code execution without login against WordPress, Drupal and Joomla.

How did AI models help find the flaws?

The flaws came to light through audits that used frontier AI models. OpenAI says Calif found a critical flaw using its GPT-5.6 Sol model under Patch the Planet, a program OpenAI presented on June 22, 2026 to support open source maintainers.

The PHP Foundation says OpenAI independently reported the same root cause and that Wiz reported separately. Wiz says it used its Wiz Atlas platform together with Google's Gemini 4 Argon and GPT-5.6 Sol, and calls Gemini 4 Argon's contribution invaluable.

The foundation itself scanned PHP with Anthropic's Claude Mythos earlier in 2026 and uses OpenAI's Daybreak Red to validate reports, though a person makes every call.

Wiz argues the models changed what counts as exploitable: chaining the bugs into an exploit was "at the edge of what human researchers could practically achieve," and the models helped it adapt exploits across applications in hours.

Why did a flaw spotted in 2024 take two years to fix?

The PHP Foundation says a core developer it contracted saw the first signs of the out of memory problem in 2024, while fixing an issue from an external audit by Quarkslab, but no concrete attack path was found and the fix stalled for months.

Each competing fix had a cost: broken backward compatibility, new denial of service risk, or changes to many third party extensions. Working by consensus and short on engineers who know the engine, the team kept stopping and restarting.

We won't pretend otherwise: the delay was a structural failure, caused by both the security team's decision process and its lack of resources.

The team is moving to majority decisions, the foundation says. AI assisted research has also swelled its inbox: 15 CVEs published since July 1, 2026, plus more than 100 reports that were invalid, duplicates or not reproducible.

How did the coordinated disclosure break down?

The coordinated disclosure held until October 20, 2026, when the CVE numbers leaked online, the PHP Foundation says, and the project pulled the release forward from October 22 to October 21.

Then Calif mistakenly posted a video of its proof of concept, whose copies sped up public exploit work, and Chainguard published patched images early, letting attackers study the fix, the foundation says. The foundation had found no functional public exploit at release time.

Cloudflare says it saw attack traffic after the leak. Side by side, the three accounts give this sequence.

Date (UTC)	Event	Source
July 30, 2026	Calif begins its PHP audit under Patch the Planet	PHP Foundation
August 11, 2026	Calif and OpenAI report the same root cause	PHP Foundation
August 19, 2026	Wiz says it found the first vulnerability	Wiz
August 24, 2026	Wiz reports to the PHP security team	PHP Foundation, Wiz
September 21, 2026	PHP acknowledges Wiz's reports and starts a coordinated fix	Wiz
October 20, about 15:00	CVE numbers leak with a wrong score of 10.0	PHP Foundation
October 20, about 20:00	Calif mistakenly publishes an exploit video	PHP Foundation
October 21, 03:02	Cloudflare firewall protections active for all customers	Cloudflare
October 21, 07:41	Chainguard publishes patched images early	PHP Foundation
October 21, 14:00	PHP releases the fixed versions	PHP Foundation

The accounts differ on one point: the foundation says the fix became its top security priority on August 11, while Wiz dates PHP's acknowledgment of its reports to September 21.

What is OpenAI proposing for future disclosures?

OpenAI on October 21, 2026 proposed a five stage framework for responding to critical flaws found with AI and asked the security community for feedback on it.

For PHP, OpenAI says it engaged the White House's Super Intelligence Force, the Office of the National Cyber Director and CISA, the U.S. cyber agency. It also published a PHP version checker as a Codex skill and says it has committed $1 billion in subsidized access to its Daybreak security tools for organizations with limited security resources.

What we don't know yet

• Whether any server was compromised through these flaws before or after the release; Cloudflare reports attack traffic, while cyber agencies in Belgium, Finland and Australia said they had seen no exploitation (The Next Web).
• How many of the other vulnerabilities OpenAI says it has found in widely used software are still unpatched, and in which projects.
• Why Chainguard published its images before the agreed release time.
• How quickly sites on PHP 8.1 and older, which get no upstream releases, will be fixed.