What Happens When AI Finds the Bug First?

On Wednesday the government of Finland switched off its own public websites on purpose (really), the Dutch national cyber-security agency took its own website offline (the agency whose entire job is cyber security, so I’m guessing they know something), Wikipedia took some of its systems down as a precaution, and the pages for Doctor Who and The Traitors went dark too (the BBC hasn’t confirmed that one was PHP, though the timing is pretty suspicious). All of them were reacting to a story the BBC broke about 2 critical flaws in PHP. If you’ve never heard of PHP, you’ve almost certainly still used it this week. It’s the programming language running underneath WordPress, Drupal, Joomla, Wikipedia, and an enormous number of websites whose owners have no idea they run PHP at all.

Those flaws had been sitting in PHP, unnoticed, for more than a decade, until researchers at Wiz (along with a firm called Calif, and apparently someone inside OpenAI around the same time) pointed frontier AI models at PHP this summer. Wiz found the first bug on August 19, could take over a Drupal site without logging in by August 20, did the same to WordPress on August 21, and reported the whole thing to the PHP security team on August 24. Wiz estimates “roughly 70%” of the internet is affected. The BBC says it can’t verify that number, and I can’t either. If Wiz is off by half, that’s still about a third of the internet.

The line from the Wiz write-up that I keep rereading is this one: “a single researcher completed in a week what previously would have taken a world-class team months of work.” I believe it. An AI agent redesigns this entire website from scratch every Saturday at 5AM (this week it’s a 1993 arcade fighting game, which tells you the agent has more free time than I do). I’d guess a human design team would want weeks for each one (I’ve never been brave enough to ask a design firm to start over every Saturday).

Fixing the PHP bugs took the rest of the world about 2 months. PHP acknowledged the reports on September 21, and the patches came out Wednesday, October 21. They even came out a little early, because somebody leaked the details on social media Tuesday and the PHP Foundation had to rush its announcement. Within hours, Cloudflare’s CTO, Dane Knecht, said they were already “seeing early attacks emerging.” The people on the other side of this have access to similar AI models.

I heard about all this from Dawn, who noticed some website was down and then saw the Wall Street Journal already writing about it (she’s the security team in our house, apparently, and the news desk). I asked one of my AI agents to read the BBC story and check every website I’m involved with. It came back with a table in about 15 minutes: roske.ai is a static site (the Saturday agent builds it, then the finished pages just sit on GitHub’s servers as plain files, with no PHP anywhere to patch), and neither the Caribbean AI Summit site nor the one server I run myself (behind a login wall, and it speaks Python) has any PHP in it either (DON’T PANIC, in large, friendly letters, at least about my stuff). (Edward, you spent 25 years running a consulting firm that installed enterprise software at Walmart, and your big security win for 2026 is that your website is too simple to hack?!? That’s like bragging that nobody can pick the lock on your tent, which, fine, technically true, but nobody asked, and you’re now explaining tent security in a post about PHP, so maybe wrap it up.)

The most useful thing in the whole BBC piece came from Don Smith at Sophos, who said finding every place PHP is running is harder than installing the patch, because it hides inside content management systems and third-party products where nobody thinks to look. So if you run a company, the question I’d ask your IT team this week is “do we know everywhere we run PHP, including inside the stuff our vendors sold us?” (The UK’s National Cyber Security Centre specifically called out appliances and services that bundle PHP, so the box in the server closet with the blinking light counts too.)

I think by October 2027 (write that down), AI will have found at least 2 more flaws in basic internet plumbing big enough that governments tell people to take systems offline, and for every one of them, the first attacks will show up within a day of the details going public. I could be wrong, because the defenders have the same AI, and they’ve organized: OpenAI started a program called Patch the Planet in June to hunt these bugs in open-source software before attackers do (Calif is part of it), and if they find the old bugs faster than the attackers do, I’ll be happy to be wrong about the count.

If you’d like me in front of your leadership team walking through what AI is doing to security (and to pretty much everything else), the speaking page has the fees and the form. Or write me (Edward@Roske.AI) and let me know how your PHP hunt went!

Asking good questions, Edward