OpenAI said on X at about 22:04 Taipei time on Oct. 21 that one of its models helped researchers identify a critical security flaw in PHP, the open-source language the company said powers about 70% of websites globally.
According to OpenAI’s website, Calif found the issue with GPT-5.6 Sol under the company’s "Patch the Planet" program. Security firm Wiz also independently discovered and reported two vulnerabilities. OpenAI said it spent about two months coordinating fixes with maintainers, researchers and security vendors, adding that the repair work was complex enough to require major changes to the PHP engine.
OpenAI described that lead time for defenders as the "Defender’s Window," a situation in which frontier AI may find a flaw before attackers do and give defenders time to prepare. The company said the case showed AI can now identify security defects with internet-scale impact.
PHP2Shell covers two RCE vulnerabilities
According to BleepingComputer, the two flaws are collectively known as "PHP2Shell" and carry identifiers CVE-2026-109618 and CVE-2026-109619. They are unauthorized remote code execution vulnerabilities affecting PHP deployments that serve web requests, including PHP-FPM and mod_php. The PHP security team said mod_php faces the highest risk.
The emergency patched versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. PHP 8.1 and older branches no longer receive official updates. Some distributions or long-term support vendors may backport fixes, but otherwise users should move to supported releases.
Wiz said the flaws sit in the PHP core engine, meaning exposure is not strictly tied to whether upper-layer applications were written securely. Common content management systems such as WordPress may also fall within the affected scope.
Cloudflare deployed protections before disclosure
Cloudflare said it had put web application firewall protections in place before coordinated disclosure and had seen early attack attempts. The UK National Cyber Security Centre called for immediate updates.
OpenAI also provided a PHP version checking tool and said it plans to release a related Codex skill. The company stressed that PHP maintainers led the patching and release process, while OpenAI, Wiz and Cloudflare contributed research, technical support and pre-disclosure defensive work.
Cloudflare said it had deployed related rules from early October and enabled protections by default for both free and paid plans about nine hours before public disclosure. Even so, it said upgrading PHP through official releases remains the permanent fix.
OpenAI launches $1 billion Daybreak support plan
OpenAI said the severity of the flaws and PHP’s broad footprint make this one of the most consequential security issues in recent years. It also acknowledged that the same level of capability could aid attackers.
To expand defensive use, the company said it will provide a total of $1 billion in subsidized Daybreak access through "Daybreak for Frontline Defenders," targeting organizations with limited resources that protect critical services.
OpenAI also laid out a five-stage response framework for high-impact vulnerabilities in the AI era: verify and patch upstream, give affected organizations time to prepare, coordinate with government partners, prepare updates across products, and publish guidance while supporting deployment. In the PHP case, OpenAI said it coordinated with the White House Super Intelligence Force, the Office of the National Cyber Director, the Cybersecurity and Infrastructure Security Agency, and international partners.
Patch adoption and exploitation remain key watch points
The case pushes AI-assisted vulnerability discovery into internet infrastructure itself. Based on the source material, the same capability can shorten the time needed to uncover complex flaws while also raising the bar for early coordination and patching on the defensive side.
For websites and Web3 services still heavily reliant on PHP and shared hosting, the immediate priority remains the same: upgrade to supported patched versions under official guidance, and treat temporary protections as a stopgap rather than a replacement. The next points to watch are patch coverage, whether large-scale exploitation appears, and whether the Daybreak subsidy program and the five-stage framework become a repeatable coordination model.