PHP Armageddon: OpenAI, Google’s Wiz find ancient holes, panicking the internet

PHP sits under roughly 70% of websites, by TechCrunch's count, including the servers behind WordPress. On Wednesday the NSA warned that two critical holes, present in every version since PHP 5, let attackers reach a server's file system or a connected database. Its advice was blunt: if a site can't be patched right away, take it offline until it can.

Patches are out from the PHP Foundation, and the recommended target is PHP 8.5.12. Anyone still on 8.1 or older is on unsupported software and faces real upgrade work, not a quick fix. Disclosure means attackers now know where to look.

This is the story we've been bracing for, and it cuts both ways. Security firm Calif, working inside OpenAI's Patch the Planet program, used GPT-5.6 Sol to find one flaw. Wiz, now part of Google, found multiple vulnerabilities with its Wiz Atlas agent, Gemini 4 Argon and GPT-5.6 Sol. Both followed responsible disclosure, worked with PHP on fixes, and pulled in Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft and Vercel to test. OpenAI says the bugs were found before hackers found them, and there were no reports of exploitation in the wild. That is a defensive win, and we should say so plainly.

Now the uncomfortable half. According to Wiz, these bugs sat for two decades partly because they were too hard to weaponize. Wiz's warning is that large language models can now adapt an exploit across PHP versions, configurations and memory layouts. The barrier that kept old bugs harmless was human effort, and that barrier is gone. The same capability that surfaced the flaw lowers the cost of using it.

So the bottleneck has moved. Finding bugs is getting cheap. Fixing them across tens of millions of servers is not. Small businesses on shared hosting, abandoned sites, appliances nobody remembers running PHP: that long tail is where the damage lands. The NSA telling site owners to unplug rather than wait is an extraordinary signal about how short the patch window may be.

I'm cautiously pro this model of AI security research, because the alternative is that someone less careful finds these bugs first. But it only works if patch distribution scales as fast as discovery. If AI keeps unearthing legacy flaws at this pace, we'll need hosting providers to push updates by default and regulators to treat end-of-life runtimes as a liability, not a choice.

WordPress said the flaw is not WordPress-specific; sites on its CMS still need their servers patched and should work with their hosts. The episode lands the same week OpenAI published a cyber-defense policy blueprint arguing that defenders are in a temporary window of advantage.