Two Serious Security Holes Found in PHP, a Tool Behind Many Websites
The U.S. National Security Agency warned on Wednesday, 21 October 2026, that two serious security flaws had been found in PHP. PHP is a programming language that runs behind many websites, doing work that visitors never see. The PHP Foundation released fixes for both flaws the same day. The NSA says any website that cannot be fixed right away should be taken offline until it can be.
Both flaws are rated 9.2 on CVSS v4.0, a scoring system where 10 is the worst. The PHP project announced the update on 20 October, according to CERT-EU. The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. TechCrunch reports that the flaws affect every version of PHP since version 5, which came out in 2004. Older versions, PHP 8.1 and earlier, no longer get support. The recommended version is 8.5.12.
The most serious flaw lets an outsider run their own commands on a server, without needing a password. That reach is large. W3Techs estimates PHP powers roughly 70% of websites, a figure TechCrunch cites. So the problem goes well beyond WordPress sites, which many people picture first. WordPress has said the flaw is not specific to its software. Site owners must update their servers and work with their hosting companies to stay protected.
The way these flaws were found is unusual. Calif, a security testing company taking part in OpenAI's Patch the Planet program, used an AI model called GPT-5.6 Sol to find one of them. OpenAI described the finding in a blog post. A Yahoo News report says Calif found the flaws in August 2026, during an audit of PHP's code that OpenAI paid for. Google's Wiz team used its own AI security tool, along with Gemini 4 Argon and GPT-5.6 Sol, to find several PHP flaws. Wiz Research says its headline flaw, called PHP2Shell, was the first PHP problem it found with that setup. A dated entry on that post, 20 August 2026, describes the flaw as one that needs no login to exploit.
Patch the Planet is an OpenAI program, run under its Daybreak initiative. It uses AI to help people who maintain free, open-source software find, check and fix bugs. OpenAI also has a separate tool called Codex Security, which it describes as an AI agent that checks for and fixes security weaknesses. That tool is in research preview, meaning it is still in early testing.
OpenAI and Wiz both say they told the PHP project about the flaws and worked on fixes before going public. OpenAI says researchers found the flaws before hackers did, and it has no reports of anyone using them yet. Its blog post calls the gap between finding a flaw and announcing it a "Defender's Window," a period when defenders can act before attackers do. Wiz thanked Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft and Vercel for helping test the findings.
Some effects are already visible. The Hacker News reports that Cloudflare saw early attempts to exploit the flaws, but no confirmed breaches. The BBC reported that websites were taken down around the world because of the disclosure.
Wiz has warned that AI language models can help attackers adjust an exploit so it works on different PHP versions and setups. Before, that adjusting was slow and costly, because many servers run different combinations. Wiz says AI removes much of that cost. Administrators should weigh this warning carefully, because the same kind of AI that found the bugs can also help someone use them.
In my view, the lesson here is about time, not the score. A 9.2 rating sounds alarming, but the practical question is how long a server has been running an old release, and whether it can be moved to a version that still gets support. Sites still on PHP 8.1 or older need a full migration, not a quick update. That work takes longer than an emergency window allows.
The broader context here is that AI tools are now finding bugs in old, widely used software. The same tools are open to people who will not report what they find. The Defender's Window only helps if maintainers can release fixes and operators can install them faster than attackers spread their methods. If that process works, the same AI tools could help remove old, hidden bugs from the infrastructure the web relies on. This episode does not yet show that, but it is why the practice of reporting flaws before going public matters.
