Hey, welcome back to Kernel Panic, the weekly newsletter from Lily Hay Newman and Matt Burgess that takes you inside the messy, fast-moving world of privacy and digital security. If you’ve been following along, you know this is normally the place where we unpack the latest breaches, weird privacy laws, and terrifying new attack techniques in a way that doesn’t make you want to throw your laptop out the window. This week, though, we have a story that feels like a science fiction premise: a set of critical vulnerabilities in PHP, the programming language that powers huge portions of the internet, suddenly exploded into public view. Security teams scrambled to patch their servers while criminal hackers tried to break in at lightning speed. Attackers launched millions of exploitation attempts within hours of the vulnerabilities being made public. And at the center of everything is an unexpected plot twist — artificial intelligence didn’t just help defenders find the holes; it also helped turn an old, neglected weakness into a real-world crisis. It’s the kind of story that makes you realize how strange and fast the security landscape has become. And if you want this kind of analysis in your inbox every week, you know the drill: sign up for the newsletter and we’ll keep you company through the chaos.
Now let’s make sure we’re all on the same page about what PHP actually is. PHP is one of those behind-the-scenes technologies that most people never think about, but it’s everywhere. It was created in the mid-1990s, back when the web was still figuring out what it wanted to be, and somehow it survived to become one of the primary languages for building dynamic websites. If you’ve ever loaded a webpage, filled out a form, logged into a blog, or bought something online, there’s a decent chance PHP had something to do with serving up that page. It’s a foundational piece of infrastructure for huge portions of the internet. That’s why a vulnerability in PHP is not just a niche technical footnote — it’s the digital equivalent of discovering that the locks on millions of apartment buildings all share the same fatal weakness. The specific family of flaws that made headlines this time got the catchy, grim nickname “PHP2Shell,” which is a blunt way of saying that an attacker could potentially go from interacting with PHP to gaining a command shell on the underlying server. That’s the prize: full control over the machine. For criminal hackers, that kind of access is gold, because it can lead to stolen data, hijacked websites, and all sorts of mayhem. And according to analysis from Cloudflare, the attacks started practically as soon as the patches arrived. Millions of exploitation attempts hit the web within hours of the vulnerabilities being published. To be fair, this is not entirely surprising: when a security patch is released, it effectively tells the whole world that a hole existed in the first place. Anyone with basic skills can read the patch notes, reverse engineer the fix, and start trying to break in. And with AI tools available to help write and adapt attack code, the barrier to entry is lower than ever.
But here’s the thing: the nightmare scenario hasn’t actually happened. There hasn’t been a mass wave of victims waking up to find their servers completely compromised, and the PHP Foundation has said it hasn’t seen any successful exploitation of the bugs. A lot of the supposedly dangerous exploit code floating around in public appears to be ineffective or not fully developed. That might sound reassuring, but it’s actually part of what makes this story so interesting. These particular vulnerabilities are not the kind of thing where you can just fire off a generic attack and magically take over every server on the internet. They’re tricky. They require a deep understanding of how PHP works, how its internals handle certain inputs, and how to piece together a chain of small weaknesses into something dangerous. In other words, they were the kind of bugs that a highly skilled human researcher could theoretically figure out, but only with a huge amount of time and effort. In fact, the underlying issues were known about, at least in part, for years. Some of the relevant flaws were identified back in 2024 but were basically set aside because they seemed too complex to exploit in the real world. Security teams looked at them, shrugged, and decided they weren’t worth the trouble. This is the part that should make you pause. As Sagi Tzadik, a security researcher at Wiz, put it: “The vulnerability had been known to the PHP team for years, but because it could not be leveraged by an attacker—at least, not a human one—it was not addressed.” Read that again. A known vulnerability sat there, unpatched, because the people who knew about it thought that no human being could actually pull off the attack. It was considered a theoretical problem, not an active threat. Then AI came along and changed the math.
That’s the real story here, and it’s worth unpacking carefully. Security researchers at a company called Calif are credited with being the first in this latest wave to identify the PHP2Shell vulnerabilities and create working exploits. They did this in early August, and they didn’t do it with AI alone. They combined expert human knowledge of PHP — its quirks, its history, its weaknesses — with frontier AI tools, including access through OpenAI’s “Patch the Planet” program. That’s the right way to think about it: AI as a force multiplier, not a magic button. The human researchers understood where to look and why; the AI was able to help them explore possibilities, test edge cases, and generate code much faster than they could on their own. What took a human team weeks or months could be compressed into days. After Calif’s discovery, researchers at OpenAI and at the Google-owned security firm Wiz independently uncovered similar PHP2Shell flaws using AI tools. So you had multiple groups of intelligent, well-resourced researchers converging on the same problem, all using advanced models to do heavy lifting. It’s a tantalizing glimpse of the future of security research. But it also points to a darker possibility. If AI can help defenders find and exploit these long-dormant bugs, it can do the same for attackers. The same technology that allowed researchers to turn a fifteen-year-old quirk into a working exploit could eventually allow a lone bad actor with minimal coding experience to do the same thing. And that means there may be more skeletons hiding in the internet’s closet, bugs that have been waiting for decades for someone — or something — smart and patient enough to finally pull them into the light.
When you step back, the bigger picture is a little overwhelming. No software is perfect. There are inevitably flaws in even the most critical and foundational digital tools, including the ones that the entire internet depends on. That has always been true. What’s changing is the speed at which those flaws can be found, understood, and weaponized. AI models are extraordinarily good at pattern recognition, at wading through massive amounts of code, and at generating the kind of brute-force experimentation that vulnerability research requires. That means the list of “too hard to exploit” bugs is getting shorter. It also means more pressure on the people whose job it is to keep the digital world from falling apart: the under-resourced IT teams, the sysadmins who are juggling too many responsibilities, the developers who have to ship features while also making sure their code isn’t a security disaster. This pressure is already showing up in data. Microsoft’s monthly security updates included a record 1,453 security vulnerabilities in a single month recently, which is a staggering number for anyone trying to keep up. And it’s not just the usual tech giants feeling the heat. Bank of America reportedly hired a thousand consultants to fix vulnerabilities discovered by Anthropic’s Mythos, an AI system used for security work. That’s an entire army of human experts brought in to clean up after an AI did the discovery. If that sounds expensive and exhausting, it is. Security researchers are quick to point out, though, that AI isn’t fundamentally introducing a new kind of threat. Instead, it’s magnifying existing dynamics that have always been part of the security landscape. Humans have been finding bugs and creating exploits for decades. AI just makes that process faster, cheaper, and more scalable. The underlying game is the same, but now it’s being played at a speed that feels almost unfair.
So what should we take away from this episode? First, a little humility: the internet runs on shoestrings and duct tape, and that has never been more obvious. The PHP2Shell saga is a reminder that some of the most critical systems we depend on are held together by code that’s older than many of the people using it. It’s not broken in the way your car might be broken; it’s broken in the sense that it was never designed for the world it now lives in. Second, AI is going to be a permanent and complicated part of the security ecosystem. It will help defenders find and fix bugs, but it will also help attackers find and exploit them. The advantage doesn’t automatically go to one side; it goes to whoever uses the tools with the most skill, care, and clear thinking. That means human judgment still matters enormously. A good security researcher brings intuition, context, and ethical judgment to the table, and no AI model can fully replace that. Third, we need to rethink how we support the people who maintain digital infrastructure. If vulnerabilities can now be found and exploited in hours or days, we can’t keep treating security as an afterthought or as a job that gets squeezed in between other responsibilities. Companies, governments, and open-source communities need to invest more in maintaining and patching the foundational software we all rely on. That’s not an exciting or glamorous conclusion, but it’s the truth. The story of PHP2Shell is not about a ridiculous new AI villain that’s going to destroy the internet. It’s about a very human problem: we built a staggeringly complex world, and we’re all still trying to figure out how to take care of it. AI just made the deadline a little tighter. Until next week, keep your patches updated, question your assumptions, and remember that the people who keep the lights on deserve your gratitude — maybe a nice coffee, too. And if you’re still not subscribed to Kernel Panic, well, now’s the perfect time to fix that.
