The age of AI-assisted offense just lapped AI-assisted defense, and your website might already be compromised.

The Summary

  • Two major security flaws discovered in PHP, the programming language running 70% of websites — both found using AI tools after hiding for over a decade
  • Government, hospital, and school sites across Europe went dark preemptively; hackers already deploying AI to exploit the vulnerabilities before patches can roll out
  • This caps weeks of AI-powered attacks hitting 100+ companies, 400 financial institutions, and critical infrastructure across six countries
  • The asymmetry is brutal: AI finds what humans missed for ten years, then weaponizes it faster than organizations can respond

The Signal

PHP powers WordPress, Wikipedia, Facebook's core systems, and roughly 70% of the web's backend infrastructure. It's the duct tape holding the internet together. Two critical vulnerabilities sat in that code for over a decade — undetected by traditional security audits, penetration testing, and bug bounty programs. Then someone pointed an AI at it.

The disclosure timeline tells you everything about where we are now. The flaws were posted to a public forum before the official announcement. The PHP maintainers scrambled to confirm and prepare patches. But by the time the coordinated disclosure happened, attackers were already probing for the weaknesses. The Finnish health agency, Port of Rotterdam, Paris police, and countless others shut down rather than risk exposure.

"PHP is one of the fundamental pieces of technology that supports the internet."

What changed: AI doesn't just automate known attack patterns. It finds novel vulnerabilities in legacy code that human reviewers have stared at for years. The asymmetry is staggering:

  • Human security researchers: expensive, slow, limited by attention span and domain expertise
  • AI vulnerability scanners: cheap, fast, tireless, improving every quarter
  • Patch deployment: still governed by human organizational speed (days to weeks)
  • Exploit deployment: automated, instant, scalable across millions of targets simultaneously

The context makes this worse. In recent weeks, AI-powered attacks hit John Deere, SharkNinja, 400 Latin American financial institutions, and governments across Ukraine, Russia, Argentina, Bangladesh, and Indonesia. This isn't a single sophisticated actor. This is commodified capability. The tools that found these PHP flaws are likely already circulating in criminal forums, getting cheaper and easier to use.

The economics broke. For years, defense had a structural advantage: one vulnerability found by a good-faith researcher could be patched before widespread exploitation. Now offense gets AI-assisted discovery, AI-assisted exploit development, and AI-assisted deployment — while defense still relies on human coordination, testing, and rollout timelines.

Here's what nobody's saying: The ten-year window matters. These weren't zero-days in new code. These were decade-old flaws in battle-tested, widely-deployed infrastructure that thousands of security professionals had examined. If AI can find those, what else is sitting in OpenSSL, the Linux kernel, or the Java runtime? The attack surface isn't just unpatched systems. It's every line of code ever written that an AI hasn't scanned yet.

The Implication

We're watching the security model break in real time. The old equilibrium assumed roughly equal access to tools and talent on both sides. That's over. AI lowered the skill floor for attackers while defense remains bottlenecked by human-speed patch deployment and testing.

If you're running any internet-facing infrastructure, your threat model just changed. Assume AI has already found vulnerabilities you don't know about. Assume attackers have weaponized them. Assume your patch window is shorter than your deployment timeline. The only edge left is reducing the time between "vulnerability discovered" and "systems hardened." That means better monitoring, faster rollout pipelines, and honest conversations about what you can't protect anymore.

Sources

The Atlantic Tech