Patch PHP Now or Pull Your Site, NSA Warns

Threat level: critical. On Wednesday the U.S. National Security Agency warned that two severe holes had been found in PHP, the server-side language behind roughly 70% of websites on the internet. According to TechCrunch AI, the flaws were found with AI tools from OpenAI and Google’s Wiz. They’ve been sitting in every version of PHP since version 5, which shipped in 2004. The PHP Foundation has already released patches. The NSA’s advice is blunt: if you can’t patch right away, take the site offline until you can.

This matters well beyond security teams. AI found these bugs, and AI is also the reason they’re now so dangerous.

🎯 Situation Report

  1. Target surface: PHP runs more than 33 million sites by some estimates. That includes the servers under popular content management systems like WordPress.
  2. Attack effect: The holes let attackers get into a server’s file system or into a database the web app uses.
  3. Age of exposure: Both flaws go back about 22 years, to PHP 5.
  4. Fix: Upgrade to the current release, PHP 8.5.12. PHP 8.1 and older versions no longer get support, so sites running them have more upgrade work ahead.
  5. Exploitation status: OpenAI says researchers found the flaws before attackers did. There are no reports of exploits in the wild so far.

🔍 How AI Found Them

Two separate teams reached the same codebase using AI.

  • Calif, a security testing firm, was working in OpenAI’s “Patch the Planet” program, which hunts for flaws in open source software. Using GPT-5.6 Sol, it found one of the flaws, OpenAI disclosed in a blog post.
  • Wiz used its Wiz Atlas security AI agent along with Gemini 4 Argon and GPT-5.6 Sol, and found multiple vulnerabilities in PHP.

Both teams followed responsible disclosure. They worked with PHP on patches before going public, then brought in others to test the fixes. Wiz thanked Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft, and Vercel. That’s a large share of the companies that host and deliver the web.

WordPress also issued a statement saying the flaw isn’t specific to WordPress. Sites running its CMS still need to patch their servers and work with their hosting companies to make sure they’re covered.

⚠️ Why the Risk Spiked

This is the key point. For two decades these bugs were too hard to exploit to be worth much to attackers. Getting an exploit to work across different PHP versions, server configurations, and memory layouts took a lot of careful manual work.

That barrier is gone.

“With AI this is no longer a barrier. LLMs can help adapt an exploit to different PHP versions, configurations and memory layouts,” Wiz warned in its blog.

So AI found and fixed this problem, and AI is also what made it high-risk. Defenders got there first this time. Now that the disclosure is public, attackers know exactly where to look, and the same models can help them turn a patch into a working attack.

📋 Tactical Orders

If you run anything on PHP, here’s your checklist:

  1. Inventory now. Find every server running PHP, including old marketing sites, staging boxes, and forgotten microsites.
  2. Check versions. Anything at 8.1 or older is unsupported and needs a real upgrade path, not just a quick patch.
  3. Upgrade to 8.5.12. Treat this as an emergency change, not a ticket for the next sprint.
  4. Call your host. If you’re on managed or shared hosting, confirm in writing that they’ve patched the underlying servers.
  5. Go dark if needed. If a site can’t be patched today, follow the NSA guidance and take it offline. A few hours of downtime costs less than a breached database.

🧭 Strategic Assessment

This is the clearest example yet of a pattern security people have been predicting. Frontier models are now good enough to find bugs that sat unnoticed in core internet infrastructure for 20 years. That’s good news for open source maintainers. It also shrinks the window between disclosure and exploitation, maybe to hours.

Expect more of these. Programs like Patch the Planet will keep surfacing old flaws in widely used software, and each disclosure will kick off a patching race like this one. Teams that can patch quickly will be fine. Teams with unpatched servers they’ve forgotten about are the ones at risk.

Full details on the disclosures and the response are in the original TechCrunch AI report.

Scroll to Top