Two critical memory-corruption vulnerabilities in PHP's core engine, collectively dubbed php2shell, were publicly disclosed on Wednesday after roughly two months of coordinated remediation involving OpenAI, Wiz Research, Cloudflare, and multiple government agencies1,2.
The flaws, tracked as CVE-2026-109618 and CVE-2026-109619, score 9.2 on the CVSS v4.0 severity scale. PHP powers roughly 70% of the world's websites, according to Wiz Research, making the attack surface unusually broad. Researchers demonstrated unauthenticated remote code execution against WordPress, Drupal, and Joomla, three of the most widely deployed content management systems.
How the bugs were found
Ryan Austin, a researcher at Calif.io, discovered one of the vulnerabilities using OpenAI's GPT-5.6 Sol as part of OpenAI's Patch the Planet initiative in August 2026. Wiz Research independently found both flaws using Wiz Atlas alongside Gemini 4 Argon and GPT-5.6 Sol.
The underlying defects are memory-safety bugs, including use-after-free and type confusion, rooted in how PHP handles out-of-memory conditions and fatal errors. When a request crashes, PHP can leave behind references to already-freed memory; the next request handled by the same worker process may treat those references as valid. Some of the vulnerable code dates back to PHP 5.4, meaning at least one flaw affects every PHP release from the past decade. PHP's engine is written in C, and Wiz Research showed that attackers could exploit the flaws to run code on WordPress, Drupal, and Joomla without logging in.
Coordinated response and patching
Four patched PHP versions were released on Wednesday, the same day the details became public: PHP 8.5.12, PHP 8.4.27, PHP 8.3.36, and PHP 8.2.35. Cloudflare began protecting its customers by default at 3:02 a.m. UTC on Wednesday, roughly nine hours ahead of the public announcement, and observed attackers modifying payloads to evade its WAF rules. Cloudflare also lists Quarkslab among the parties that disclosed the bugs.
OpenAI described the effort as part of Patch the Planet and said its work involved the Cybersecurity and Infrastructure Security Agency, the Office of the National Cyber Director, and the Super Intelligence Force. OpenAI also proposed a five-stage framework for handling critical vulnerabilities that frontier AI models help uncover.
ANALYSIS The company released a PHP version checker with two Codex skills for checking local servers and websites, though the tool can flag vulnerable versions but cannot confirm whether a patch has been applied, particularly when vendors backport fixes.
OpenAI's broader cybersecurity push includes Daybreak for Frontline Defenders, a $1 billion subsidy for access to Daybreak cyber models and products introduced roughly two months ago.
ANALYSIS The parallel, independent discovery by Calif.io and Wiz Research using different frontier models demonstrates that AI-assisted vulnerability research is producing results across multiple teams simultaneously, compressing the window between a bug's existence and its detection.