Home › AI news

AI found a 14-year-old hole in 70% of the web. Here's what actually happened.

Two critical holes in PHP, the software behind most websites, were found with AI's help. Then the secret leaked early. Here's what PHP2Shell is, who's actually been hacked, and what you should do.

Illustration: a cracked padlock made of glowing code on a website storefront, with a small robot patching the crack and a shadowy pixel figure reaching for it, under a night sky of network lines and server racks.
Illustration made with Runway AI for Just Spilling the Bits.

If a bank, airline or government website went dark on you this week, this is probably why.

Two serious security holes turned up in PHP, the behind-the-scenes software that runs roughly 70% of websites whose server language can be identified. That includes most WordPress, Drupal and Joomla sites. Security people nicknamed the pair PHP2Shell. The twist that makes it an AI story: AI helped find them.

What PHP2Shell is, in plain English

PHP is the engine many websites run on. These two flaws (CVE-2026-109618 and CVE-2026-109619) are in the engine itself, not in any one website's code. That means a site can be perfectly well kept, with every plugin updated, and still be exposed if the PHP underneath is old.

Worst of all, an attacker doesn't need to log in. A carefully built web request can let them run their own commands on the server and grab passwords and data stored there. PHP's own team rated both flaws 9.2 out of 10 for severity, and researchers say the oldest one goes back to PHP 5.4, released in 2012.

PHP shipped fixes on Oct. 21: versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. PHP 8.1 and older are past end of life and get no official update, only bare-bones patches that admins have to apply by hand.

Where AI comes in

On Aug. 1, researcher Ryan Austin at the security firm Calif found the first bug with help from OpenAI's GPT-5.6 Sol model, as part of OpenAI's "Patch the Planet" program. Weeks later, on Aug. 19, the cloud security firm Wiz found the same problems on its own, using Google's Gemini 4 Argon, GPT-5.6 Sol and its in-house Wiz Atlas tool. Wiz says one researcher did in about a week what would have taken a top team months.

One correction to the hype: AI didn't find this alone. Human researchers pointed the AI where to look and checked its work, and PHP's own volunteers and staff wrote and shipped the fix. AI was the power tool, not the carpenter.

The rough part: the secret got out early

The plan was a quiet, coordinated release. Instead, details leaked online on Oct. 20, a day before the fix. Amazon later said it had spotted an exploit attempt as early as Oct. 17. For a short window, attackers knew about the hole before anyone could patch it.

So websites around the world did the safe thing and unplugged:

  • Europe: government and university sites in the Netherlands, Finland and Austria went offline ahead of the patch.
  • Japan: Prime Minister Sanae Takaichi spoke about it publicly, and services from childcare systems to online games paused.
  • The web: the Internet Archive briefly shut down to patch.

Going offline doesn't mean getting hacked. Most of these were precautions, the website version of locking the doors while the locksmith works.

So has anyone actually been hacked?

Until this weekend, the honest answer was "no confirmed break-ins." That has changed.

Cloudflare says it saw millions of exploit attempts after the flaws went public. Then, in a report published today, Palo Alto Networks' Unit 42 said attackers broke into a European government agency through PHP2Shell and, within two hours, reached password data and the secret codes behind two-factor logins. Unit 42 also says the attackers' commands looked like they were run by an AI agent working through the break-in step by step.

Separately, the security firm CYFIRMA says Korean-speaking hackers, possibly tied to North Korea's Lazarus Group (CYFIRMA rates its confidence only "medium"), have lined up 216 unpatched servers at governments, banks and airlines to hit next.

And PHP has flagged a third, related flaw (CVE-2026-115560). It only bites with an unusual setup, and the fix will come in PHP's next regular release.

What you should do

If you just use the internet:

  1. Be patient with outages. If a bank or store site is down this week, it's almost certainly patching, not a hack. Give it a few hours.
  2. Watch for odd behavior on familiar sites: surprise downloads, strange redirects, or being asked to log in over and over. Close the tab if something feels off.
  3. Keep your phone, browser and computer updated. This flaw lives on servers, not your phone, but updates protect you from everything else going around.

If you run a website or home server:

  1. Check your version. Run php -v, then update to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 and restart PHP.
  2. On Ubuntu or Debian, the version number can look old even when your package is patched. Check the package changelog for the two CVE numbers.
  3. On shared or managed hosting, ask your host whether it's patched. WordPress.com, WordPress VIP and Pressable say their customers are already covered.
  4. On PHP 8.1 or older, upgrade. There's no official fix coming.
  5. Can't patch today? Cloudflare's free plan includes firewall rules that block known attacks. That's a stopgap, not a fix.

My take: This story cuts both ways, and that's exactly why it matters. The good news is real: AI gave the good guys about a two-and-a-half-month head start on a bug that sat unnoticed for over a decade. Without it, criminals might have found it first. The bad news is just as real. The same kind of AI that found the hole is now helping attackers work through break-ins in hours instead of days, and the leak shrank the head start to almost nothing. And here's the part that bugs me: PHP runs most of the web, but it's kept going by a small foundation that took in about $730,000 in donations last year. Some AI labs spend that before lunch. If AI is going to keep turning up old holes this fast, the people who have to fix them need more help than a tip jar. Is my own site safe? Yes. Just Spilling the Bits runs on Cloudflare Workers, which doesn't use PHP at all.

Coming up next

Humanoid robots that fold your laundry and load the dishwasher: what's real, what's hype, and when you could actually buy one. That's the next Science Bits, with a video.

Everything here is written in my own words from the sources below. I used AI (Claude, by Anthropic) to help research and draft this, then checked the facts against these sources. OpenAI and Google, both mentioned above, make AI models that compete with Claude. This story is still developing, and details may change.

Sources:

Hi, I'm Thomas

I've worked with my hands for 20+ years in industrial construction, electrical, automation and refrigeration. No programming background. I test AI tools after hours and tell you what actually works.

The site is free to read, and always will be.

Read my full bio →