OpenAI's Patch the Planet: AI finds critical PHP flaws

What happened

On Wednesday OpenAI described (OpenAI) how it spent two months coordinating an international response to two critical PHP vulnerabilities. Security firm Calif found one of the flaws using GPT-5.6 Sol as part of OpenAI's Patch the Planet initiative. Wiz independently discovered and reported both. PHP's maintainers led the fixes, with Wiz, Cloudflare and other partners contributing research, technical support and coordination.

OpenAI calls it one of the most impactful security issues in recent memory. To its knowledge, frontier AI found the bugs before attackers did, opening a “Defender's Window” to prepare fixes and protections ahead of disclosure.

The two PHP flaws

The PHP team released patched versions on October 21 (GitHub). According to Help Net Security, the flaws have been dubbed PHP2Shell:

• CVE-2026-109618: remote code execution via out-of-memory (OOM) corruption. Internal request data can be left partially initialized, and request cleanup then treats it as valid.
• CVE-2026-109619: remote code execution via stale global state. Request-owned internal state can outlive its freed memory, and a later request on the same worker may use it.

Patched versions: PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Help Net Security reports that every PHP installation serving web requests is affected (PHP-FPM, mod_php, php-cgi, FrankenPHP, LiteSpeed), with mod_php setups at greatest risk. PHP 8.1 and earlier get no official release, but the PHP security team published minimal source backports for 7.4, 8.0 and 8.1. Cloudflare says sites behind it already have WAF rules for the known exploit vectors, including on the free plan.

A five-stage response

OpenAI also proposes a framework for responding to critical flaws in widely used software in the AI era. It describes five stages:

• Validate and fix the upstream vulnerability, including adversarial testing with frontier models to look for patch bypasses and variants.
• Give affected organizations time to act. Distributors, hosting and security providers can prepare in advance, for example with WAF rules. Maintainers decide whom to involve.
• Coordinate with government partners. For PHP this included the US ONCD, CISA and the White House's Super Intelligence Force, as well as international partners.
• Prepare updates across affected products.
• Publish guidance and support deployment once a patch is available.

OpenAI is asking the security community for input. How long should disclosure windows be? Who needs early notice? Should governments ever use a vulnerability to patch systems they own after losing administrative access? Feedback is open to the public.

Money and tools for defenders

OpenAI has committed $1 billion in subsidized Daybreak access through Daybreak for Frontline Defenders, aimed at organizations with limited security resources that protect essential services. Open-source maintainers get subscriptions through Codex for OSS, and OpenAI also sponsors projects such as the Rust Foundation. It has released a PHP version checker, with a Codex skill to identify vulnerable versions coming soon.

OpenAI acknowledges the dual-use problem: the same capability that finds bugs can help build exploits. It is calibrating additional blocking classifiers for Daybreak while trying not to block legitimate security work.

What it means

If you run PHP, patch first. Inventory every PHP install, including old and forgotten apps. The bigger picture: OpenAI expects AI-found vulnerabilities like these to rise sharply in the coming months. If you maintain an open-source project, Patch the Planet's support program is worth a look before the next disclosure is about your code.