The U.S. National Security Agency issued an urgent warning on Wednesday about two critical vulnerabilities in PHP, the server-side language that underpins roughly 70% of the web — by some counts more than 33 million sites. Patches have already been released by the PHP Foundation, but the NSA's guidance is blunt: if a site cannot be patched immediately, it should be taken offline until it can be. The flaws allow attackers to reach into a server's file system or the databases behind a web application.

Decades in the Code

What makes the disclosure especially alarming is the age of the defects. The vulnerabilities have existed in every version of PHP since 5, released in 2004, meaning they have sat unnoticed in a vast swath of production infrastructure for more than two decades. The fix path is not uniform: PHP 8.1 and older branches are no longer supported, so operators on legacy versions face an upgrade rather than a simple patch. The recommendation is to move to the current release, PHP 8.5.12. That is a significant undertaking for organisations running long-lived applications, custom extensions or tightly pinned dependencies.

The discovery itself came from AI-assisted security research. Calif, a security testing company engaged in OpenAI's "Patch the Planet" initiative to hunt for flaws in open-source software, found one of the vulnerabilities using GPT‑5.6 Sol, OpenAI disclosed in a blog post. Separately, Google's Wiz team used its Wiz Atlas security AI agent alongside Gemini 4 Argon and GPT-5.6 Sol to surface multiple vulnerabilities in PHP. Both teams followed responsible disclosure practice, working with the PHP Foundation on patches before going public, then collaborating to validate the fixes. Wiz credited Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft and Vercel for their involvement in that effort.

Why AI Changes the Risk Calculus

The timing matters as much as the technical details. OpenAI says the flaws were found by researchers before attackers discovered them, and there have been no reports of exploitation in the wild. Part of the reason is historical: before modern AI tooling, these bugs were difficult for attackers to weaponise reliably. Wiz warned in its own blog that this barrier has now fallen. Large language models can help adapt an exploit to different PHP versions, configurations and memory layouts, turning what was once a specialist undertaking into something far more repeatable. The same capability that allowed defenders to find the holes at scale also lowers the cost of attacking them.

That dynamic has turned disclosure into a race. Now that the flaws are public, the window between awareness and exploitation is measured in days or hours, and the internet must patch tens of millions of sites before unpatched ones are targeted. The NSA's advice to pull vulnerable sites offline reflects the reality that patching at this scale is not instantaneous. Web hosting providers and CDN operators — several of whom participated in the patch validation — are likely to be central to remediation, since many sites depend on upstream platform updates rather than direct server administration.

WordPress and the Hosting Layer

PHP's reach extends through popular content management systems, most notably WordPress. WordPress has issued a statement clarifying that the flaw is not specific to its CMS and that sites running it must patch their servers and coordinate with their hosting providers to ensure protection. That distinction matters for the many site owners who manage content through an administrative interface but do not control the underlying runtime. For them, the practical question is whether their host has updated PHP and whether any custom code depends on behaviour that changed.

For developers and IT teams, the immediate priorities are version inventory and dependency checks. Any service still on PHP 8.1 or earlier needs a migration plan, and teams should verify that their frameworks, extensions and third-party libraries are compatible with 8.5.12 before rolling it out. Where an upgrade cannot happen quickly, isolating or temporarily removing the affected service is the safer option the NSA recommends. The episode also serves as a case study in AI-driven vulnerability research: models are now capable of finding deep, long-lived defects in critical open-source infrastructure, which means the same techniques will likely be applied to other foundational projects — and defenders should expect the disclosure tempo to accelerate accordingly.