AI finds decade-old PHP flaws: 75% of the world's websites at risk, government sites pulled offline, 1.5 million hosts still unpatched

The internet is going through the AI era's first infrastructure-grade security crisis: on Wednesday the NSA issued an emergency warning that researchers using AI tools had found two critical vulnerabilities in PHP — the language underpinning about 75% of the world's websites — where the flaws had lurked for more than a decade; government websites in the Netherlands, France, Denmark and Finland were taken offline, and the PHP team shipped an emergency patch urging all users to "update immediately"^[1].

The picture: AI found the bugs before attackers did

The timeline itself is historic. Google-owned security firm Wiz disclosed in a blog post that its autonomous AI vulnerability-research tool, Wiz Atlas, found the flaws using Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol; another security firm, Calif, said it independently found the same vulnerability on August 1 using large language models and has since helped with remediation, with the two firms coordinating with the PHP team and OpenAI^[1]. Wiz noted the flaws had hidden in the PHP codebase for over a decade and would have been extremely hard to find — let alone exploit — by manual review: "before AI, this manual work was hard to do and harder to scale; with AI, that barrier no longer exists"^[1]. OpenAI's statement maxed out the severity scale: "given the severity of these vulnerabilities and PHP's wide usage, we believe this is one of the farthest-reaching security issues in recent memory," adding that "to our knowledge, frontier AI found these vulnerabilities before any attackers"^[1].

Quantifying the risk: 2.6 million hosts, 1.5 million unpatched

The numbers sketch the attack surface. Per internet-intelligence firm Censys, of roughly 2.6 million hosts running PHP, about 1.5 million currently run versions with no patch available; PHP underpins about 75% of websites with a known programming language, with WordPress, Drupal, Joomla and Magento all built on it — Drupal Association interim CEO Tiffany Farriss's judgment: "every server running any PHP version from the past 20 years is at risk; this affects the majority of sites on the internet"^[1]. More urgent is the post-patch reverse-engineering race: with patches now public, experts warn hackers can use AI to easily reverse-engineer the vulnerabilities, and a mass-intrusion wave could be one trigger away — TLP:Black co-founder Costin Raiu calls it "a race to exploit first," with the next step "mass compromise of every vulnerable web server on the internet"^[1]. The industry’s "PHP Armageddon" framing says something in itself — the panic has become part of the story’s reach^[2]. The risk distribution is highly uneven: Automattic, which operates WordPress.com, says its hosted users and Tumblr users are unaffected — the real soft underbelly is small, resource-strapped PHP sites, the long tail that zombie networks love to prey on^[1].

The structural shift: AI accelerates both offense and defense

The deeper significance is a structural shift in the offense-defense balance. On defense, AI has turned needle-in-haystack code auditing into a scalable assembly line — flaws hidden for a decade were located by AI first, the most convincing field proof yet of "AI for Security." But the coin's other side: attackers gained the same scalability — patch reverse-engineering, exploit generation, attack-chain assembly, every link accelerated by AI. In other words, AI hasn't made the world safer or more dangerous; it has sped up both sides' clocks — and offense and defense are inherently asymmetric: defenders must patch 2.6 million hosts, attackers need only find one unpatched box. For site operators the checklist is short but hard: upgrade PHP to the latest patched version immediately; for legacy systems that can't take the patch, consider virtual patching at the gateway; users on major platforms should confirm platform advisories; and dependencies and plugins across PHP applications need a synchronized check.

What to watch

First, the scale of exploitation attempts in the next 72 hours — honeypot data from security vendors will give the first answer. Second, how fast the 1.5 million unpatched hosts converge — deciding whether this is a contained incident or a slow-burn epidemic. Third, the next step in AI vulnerability research — whether Wiz Atlas-style tools spread across vendors and open-source communities into a new norm of "AI-inspected codebases." Fourth, the policy reaction — the NSA has moved; whether EU cyber agencies and national CERTs' coordination spawns mandatory audits of critical open-source components^[1].