OpenAI spent two months coordinating the repair of two critical vulnerabilities in PHP, the open-source programming language behind a great many websites and web applications. In its Patch the Planet post, OpenAI says that to its knowledge an AI model found the flaws before attackers did, which gave maintainers and security companies time to prepare fixes before anything went public.
The discovery came from two directions. Calif, working under the Patch the Planet initiative, spotted a critical flaw with the help of GPT-5.6 Sol, OpenAI’s model, and Wiz independently discovered and reported both vulnerabilities. PHP’s maintainers wrote and released the fixes. Wiz, Cloudflare and other partners supplied research, technical support, and coordination.
The post does not name the flaws, give identifiers, or say how many sites are exposed. It says only that PHP is widely used and that OpenAI ranks this among the most impactful security problems it can remember. That is the company’s own judgment, and readers have no independent measure of it yet.
For anyone running PHP, OpenAI’s advice is to apply the fixes. To help, the company released a checker for PHP versions, packaged as a Codex skill, which flags potentially vulnerable PHP versions so operators can work out which updates apply to them.
The larger piece is a proposed framework for the next time an AI turns up a flaw in software that many people depend on. It has five stages. The first is to validate and fix the problem upstream, including attempts by frontier models to break the patch and find related variants. The rest follow in order:
- Buy affected organizations time to act, such as distributors, hosting providers, and security vendors preparing firewall rules.
- Bring in government partners to protect critical systems.
- Prepare updates across affected products.
- Publish guidance once a patch exists, and support deployment.
For PHP, the government outreach covered several parts of the US government: a White House unit called the Super Intelligence Force, the national cyber director’s office, and CISA (the federal cyber defense agency), plus international partners. Maintainers, not OpenAI, decide who gets advance notice, and the timeline gets reassessed if attackers are already exploiting the bug, details leak, or a public patch gives it away.
OpenAI also put money behind the defenders. It has committed $1 billion in subsidized access to Daybreak, its security tooling, through a program called Daybreak for Frontline Defenders, aimed at organizations with limited security resources that protect essential services. It offers Codex for OSS subscriptions to all maintainers and sponsors the Rust Foundation directly. It is also tuning extra blocking classifiers for Daybreak, and concedes that safeguards against abuse can block legitimate security work.
Notice who is writing the rules here. OpenAI builds the models that find these bugs, funds the people who fix them, and is now drafting the process for disclosure, while asking the security community to critique it. Its open questions are sharp ones: how long disclosure windows should be, and when governments might patch systems they own or manage after losing administrative access. The post even imagines internet-wide scanning for known vulnerabilities that either notifies site owners or patches the system.
OpenAI expects findings like these to rise sharply over the next few months, so PHP will not be the last emergency of this kind. Anyone who runs PHP should run the checker this week and treat shorter patch windows as the new normal.
Reported by OpenAI on 21 October 2026.
