Every PHP server from the last decade just got owned... Transcript, AI Summary & Key Points

00:00 It feels like it was just yesterday when the biggest risk to your infrastructure was the jaded Silicon Valley developer going manic and deleting his package from the mpm registry. But nowadays, it's a bit more bleak. In just the last 72 hours, the literal NSA encouraged anyone using PHP to go dark. A Python package with 23 million weekly downloads got backed.

00:19 A GitLab found a bug that's been affecting every release since 2014. The US banks are detecting unauthorized thirdparty agents operating inside customer accounts. and an Indian certificate authority that issues HTTPS certificates for the entire internet was generating private keys with a 2D 100. In today's video, we'll break down the most vulnerable week in the history of the internet and once again look at how to update your infrastructure fast enough to not get zero day but slow enough to not get supply chain attack.

00:47 It is October 23rd, 2026 and you're watching the code report. The PHP stands for personal homepage and was created in 1994 by a Danish guy named Rasmus Lordorf who admitted he actually hates programming just like me. Since then, PHP has been co-opted by the elite who love to flaunt their lavish lifestyles with no regard for the workingclass developer.

01:07 Well, this week the joke's on them because PHP finally did some wealth distribution of its own when two bugs now known as PHP2 shell let anyone with an internet connection run whatever they want on any PHP server released in the last decade by stringing together two flaws in how PHP manages memory. The first one comes from the fact that every PHP server has a memory limit and when a request goes over it, it uses an old C trick to jump straight to the end of the request without finishing or cleaning up anything it was

01:35 working on. that the second one comes from the fact PHP recycles the same process for every request and the last one left in memory doesn't always get cleared before the next one arrives. For 20 years, neither of those was a problem. But recently, researchers figured out that if you can bind them by pushing PHP over its memory limit at exactly the right moment, like halfway through a file upload, it'll leave behind halfbaked data that the next request can pick up and use, which if exploited means anyone can run their

02:01 own code on your server. And this vulnerability was actually found a few months ago by a security firm using GPT 5.6 Soul who brought it up with the PHP team who then spent the last few months working on a fix that was supposed to drop yesterday. But that plan fell apart on Monday when the bugs ID leaked on Hacker News. And then the same firm who found it accidentally posted a video of themselves exploiting it on Tuesday.

02:22 So that left 24 hours where the exploit was public, but the patch wasn't. The ramifications of which I'm sure we'll be hearing about for years to come. But that was only the beginning. On Wednesday, Node.js shipped a fix for a bug that lets an attacker run code on your server by sending it a malformed compressed file. And GitLab fixed one in the Regax engine it bundles from Ruby that let anyone run code on a self-hosted server that's been there since 2014.

02:46 Thankfully, if you just stay up to date on your dependencies, you'll be fine. Unless one of them is a Python library that most AI apps use to talk to models called light LLM, where on Wednesday a new version shipped that stole any important credential on your machine and was up for a couple of hours before Pi pulled it. But at least all of those hacks required some state-of-the-art models stringing together esoteric vulnerabilities.

03:07 In this next one, it was just pure organic human incompetence. Every time your browser shows you an HTTPS lock, it's because a certificate authority confirmed that the website you're looking at is who it claims to be. that that whole system depends on each authority keeping a handful of private 384-bit randomly generated keys secret. But last Saturday, a researcher noticed that eight of the keys belonging to an Indian certificate authority that's trusted by every major browser were integers smaller than 256, which in

03:38 practical terms meant that anyone could start impersonating any website on the internet by simply guessing a number between 1 and 255 and then issuing a valid HTTPS certificate. Immudra, the company responsible for the incompetence, blamed a bug in their key generation script. And I blame them for thinking that a 384bit meant that you could pick a number between 1 and 384.

03:59 Chrome, Firefox, Apple, and Microsoft all blocked the certificates within 3 days of discovering the issue. And Mozilla pulled one of Emra's root certificates out of Firefox entirely. And to finish off the week, Morgan Stanley, JP Morgan, and Bank of America say they've been finding agents logged into customer accounts doing things the customer never asked for.

04:17 And this tracks because OpenAI recently disclosed that two of its own coding agents decided it was morbid time, and turned off alerts to its security team, removed their own monitoring, and when one was asked about it, it acted like nothing happened. >> And all of this was sandwiched between two breaches that were both targeting the same thing. Last Friday, someone broke into Open Code looking for users OpenAI and anthropic API keys, which Open Code says was part of a large-scale distillation attack.

04:43 And this week, Railway had to pause every customer deployment after attackers decrypted some customers environment variables, which Railway says was also part of a broader campaign targeting AI inference providers, which in a way is kind of a relief because no one wants your data anymore, just your API keys. But if you want help protecting your own projects, check out Smrep, the sponsor of today's video.

05:04 Everyone is slapping the words Mythos class on their security tools right now. So, Sgre actually ran Mythos, GPT Soul, and a bunch of other Frontier models through their own vulnerability benchmarks. And they found that the harness makes a much bigger difference than the model. The SEM Greps harness correctly identified 6.5 times as many real security bugs as the same model on its own because it gives the LLM real program analysis with data flow and dependency context.

05:29 And their guardian tool puts that inside your coding agent to protect it from itself. One command lets you install it as hooks and an MCP server. And every file your agent writes gets scanned right after it's written. Like here, Claude Code just tried to hardcode my live stripe key, but Guardian blocked it before it touched git and the agent fixed it on its own.

05:49 And it can do the same for vulnerable patterns and malicious packages right as they happen. But try Guardian for free for 60 days using the link below, which is double their normal trial. This has been the code report. Thanks for watching and I will see you in the next one.