In partnership with
What's Actually Happening
This week, governments in Finland, France, Denmark and the Netherlands took their own websites offline on purpose. So did Japanese airlines and banks, an Irish news site, and parts of Wikipedia.
The cause was PHP2Shell, a pair of critical bugs in PHP, the programming language behind roughly 7 in 10 websites whose server language is known. It lets an attacker take over a web server without logging in. The bug had been hidden in PHP's code since 2012.
The twist is who found it. Security teams used AI to find the flaw and then turn it into a working attack. One firm says a single researcher did in a week what used to take a world-class team months.
So AI found it first, which is the good news. The bad news is that it wasn't by much.
ARTIFICIAL INTELLIGENCE
🐘 What PHP2Shell Actually Is
Two bugs, both rated critical (9.2 out of 10):
🧠 CVE-2026-109618: when PHP runs out of memory at the wrong moment, it can leave internal data half-built and then treat that garbage as valid.
👻 CVE-2026-109619: data from one web request can survive after its memory is freed, and the next request on the same server can pick it up.
Chain them together and an attacker can run their own code on your server, often before your app's code even starts running. Security researchers say they confirmed pre-login takeovers against specific versions of WordPress, Drupal, Joomla and Magento.
Who's exposed: basically every PHP setup that serves web traffic, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed. Apache with mod_php is the riskiest.
The fix: update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35. Versions 8.1 and older don't get official releases. The PHP team posted minimal source patches for 7.4 through 8.1, and anything older has no fix at all.
🤖 How AI Found It
Three teams found the bugs in late August, and all of them used AI:
🔎 Calif, a security research firm, found a critical flaw using OpenAI's GPT-5.6 Sol through OpenAI's Patch the Planet initiative. Its researcher published the technical writeup.
🧪 Wiz independently found both bugs. It compares the severity to Heartbleed, the 2014 bug that shook the internet.
🏛️ OpenAI then coordinated a two-month private fix with PHP's maintainers, Cloudflare, CISA and international governments.
Wiz described what changed: "What AI changed was the ability to understand the weaknesses and turn them into a working exploit." Finding a bug used to be the hard part. Now AI can find the bug and write the attack too.
OpenAI says this is "one of the most impactful security issues in recent memory," and that it expects findings like this to "increase sharply in the coming months." It's committing $1 billion in subsidized access to Daybreak, its AI defense program, for under-resourced groups that protect essential services.
12 ways to build real wealth
Most people have exactly one income stream: their job. Marina Mogilko (AKA Silicon Valley Girl) has twelve.
The Hustle's free wealth-building guide reverse-engineers how she got there by turning side projects into a 7-figure portfolio of courses, products, and passive commissions. It's a practical blueprint you can borrow one piece at a time.
Get it for free when you subscribe to The Hustle and join 1.5M+ readers.
⏱️ The Four-Day Problem
OpenAI says frontier AI found these bugs "before attackers did." That's probably true. But it wasn't before attackers found out.
The timeline:
🔒 Late August: bugs found, and the fix starts in private.
⚠️ October 17: Amazon's honeypot network catches a PHP2Shell exploit attempt, four days before anyone outside the coalition was supposed to know.
📣 October 20: PHP pre-announces a critical patch coming "within 24 hours."
🩹 October 21: patches go out. Within hours, proof-of-concept exploits are circulating and multiple attackers are probing servers.
Nobody knows how the October 17 attacker found out. Amazon says every attack it saw was an attempt, with no confirmed successful break-ins so far, but it's advising companies to start their investigations from October 17, not from the disclosure date.
Chief Information Security Officer @ Anthropic
Amazon's line sums up the whole situation: "An embargo protects defenders only until someone else finds the same bug."
🌍 The Internet Went Dark (On Purpose)
The FBI told organizations that can't patch immediately to take their sites offline, and many did:
🇫🇮🇫🇷🇩🇰🇳🇱 Government websites in Finland, France, Denmark and the Netherlands, including the Dutch cybersecurity agency itself.
🇯🇵 ANA, JR West and MUFG in Japan suspended pages or services. Japan's prime minister told companies to consider doing the same.
📰 The Journal, Ireland's second most-used news site, went offline.
📚 Wikimedia took some systems down as a precaution.
🙋 What This Means for Regular People
If you're not running a website, here's the short version from experts speaking to CBS:
✅ Visiting a PHP site doesn't infect your device.
⚠️ The risk is your data on someone else's server: your name, email, address and birthday stored on a site that gets breached.
🎣 Watch for phishing. Nobody knows yet what was taken, if anything, so be extra careful with unexpected emails and texts this week.
🛒 No need to stop shopping or browsing.
🔓 Why It Matters
If you run anything on PHP: update tonight. On a VPS, run php -v. On managed WordPress or shared hosting, check that your host has patched. Then review your logs back to October 17.
For the industry: this is the biggest bug yet found with AI doing the heavy lifting. Two days ago we wrote about one hacker using AI agents to breach 100+ companies for about $50 each. PHP2Shell is the same trend from the other direction. AI defenders found it, AI-assisted attackers caught up within days, and the gap between those two moments is now the whole fight.
For your patching habits: Amazon's advice to security teams is to measure patch time in hours, not the usual 7 or 30 days. If a bug can go from discovered to exploited in a week, a monthly patch cycle is too slow.
Top 5 In AI Research 🔬
The stories moving fast beyond today's headlines:
🎓 Perplexity's JOLT trains one model as both teacher and student: the teacher learns to give feedback the student can actually use, with reported gains in math, coding and agent tasks.
☁️ OpenAI's models are coming to Google Cloud: OpenAI frontier models and Codex will appear in Google Cloud's Model Garden, with general availability in the coming weeks.
🏭 Mistral and NVIDIA team up on AI for industrial engineering: including public datasets and benchmarks for agents doing engineering and chip design.
🪖 Google DeepMind staff resign over Pentagon AI oversight: Google says its public-sector deals go through compliance reviews.
🧭 Dario Amodei defends Anthropic's "race to the top": he points to safety calls that cost Anthropic commercially, including a four-month delay on its Mythos model.
🛠️ Tools That Are Hot Right Now!
🪨 Flint: Recursive's local coding model, built by an automated AI research system, with about 2.5B active parameters per token.
🗣️ HeyGen Voice Turbo: now #1 on Artificial Analysis' controlled-voice leaderboard, at $4 per million characters.
📐 UseIt Studio: a design agent that outputs editable, print-ready manufacturing files with dimensions, bleed and dielines.
What's The Recap?
For decades, the internet's security rested on a quiet assumption: finding a deep bug in widely used code took rare skill and a lot of time, so most bugs stayed buried. PHP2Shell sat in plain sight in code that runs most of the web for 14 years, until AI read it. That's the real story. AI isn't creating new weaknesses. It's revealing ones that were always there, much faster than humans can fix them. OpenAI expects many more findings like this, which means the next few years will bring lots of "patch tonight" moments, and the winners will be the teams that can patch in hours. If you take one thing away, make it this: the time between "a bug exists" and "someone is exploiting it" is now measured in days, so your updates need to move just as fast. Start with php -v.
What Do You Rate Today's Newsletter?
Stay building. 🤖