On 21 October, the PHP project did something it almost never does: it shipped emergency security fixes. The reason was a pair of flaws, tracked as CVE-2026-109618 and CVE-2026-109619 — memory-safety bugs rated as high as 9.2 out of 10 on the industry severity scale, each capable of remote code execution, and affecting nearly every PHP release going back to version 5.4, according to Wiz, the cloud-security company that published its technical disclosure this week. The fixed release is PHP 8.5.12, and the US National Security Agency's advice, relayed by TechCrunch, is unusually blunt: if you cannot patch, take the server offline. PHP powers roughly 70% of the world's websites — the quiet plumbing underneath millions of small-business sites, school portals and online shops.
What makes this incident different from the last decade of PHP patches is who found the flaws. Wiz says its Wiz Atlas agent found them with the help of two frontier AI models — Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol. One of the two flaws was found independently a second time, by the security-testing firm Calif, using GPT-5.6 Sol inside OpenAI's 'Patch the Planet' programme. This is the first incident at this scale in which AI systems were the vulnerability hunters rather than the subject of the vulnerability.
The speed of the hunt is the story's uncomfortable half. Details of the flaws leaked before the planned coordinated disclosure, and security firms reported exploitation attempts — some of them before the fixes had even shipped, according to SecurityWeek. The patches had to race the exploits. Wiz's own warning is the sentence worth quoting honestly: what kept old bugs from being farmed at scale was the tedium of adapting an exploit across versions, configurations and memory layouts, and AI models erase that. The same technology that found these flaws also lowers the cost of weaponising them.
The other side of the coin arrived the same week. On 22 October, Microsoft attributed a supply-chain compromise of the IT management vendor ConnectWise to a China-aligned group it calls Storm-0940, running an open-source penetration-testing framework called ARTEX that, in Microsoft's words, executes 'an autonomous penetration test from reconnaissance through data exfiltration' with little human involvement. Microsoft says it logged sign-in attempts against more than 2,250 tenants, peaking at 2,900 a minute, with some attacks reaching privileged cloud access within three hours. Attackers and defenders are now running the same class of software. That is the new equilibrium: AI on both sides of the wall.
What does this mean at ground level, for people who are not security engineers? If you run a small business — a clinic, a school, an online shop — there is a good chance your website runs on PHP, most likely through WordPress, which still powers a huge share of small-business sites everywhere. Your exposure does not depend on your size; it depends on whether anyone keeps the software under your site updated. The most useful paragraph in this story is the boring one: know what your site runs on, switch on automatic updates where they exist, and keep a backup you have actually tested. The 'we will update it next month' habit is now, measurably, the vulnerability.
For students, the lesson is equally direct, and far more encouraging. The work Wiz and Calif did this week — reading code at machine scale, reasoning about memory safety, proving a flaw — is exactly the kind of work AI-assisted security professionals now do, and it is learnable work. Reading a CVE advisory, checking a server's version, applying a patch and verifying it afterwards: that sequence is an employable skill in Karachi as much as in Berlin. The barrier to entering this work has never been intelligence. It has been exposure and practice.
Where does AstolixGen fit into this, honestly? We are not a security company, and a blog article does not make us one. What we do build — automation and AI agents for small businesses — sits one step away from the real lesson here: most of what harms small organisations online is not cinematic. It is failed maintenance — software never updated, dependencies nobody knows about, backups nobody tested. That is a workflow problem before it is a technology problem, and workflow problems are the ones automation is genuinely good at: scheduled checks, alerts when something drifts, reports a non-technical owner can actually read. We build that kind of plumbing. This week's news is a reminder of why unglamorous plumbing matters more than shiny tools.
The honest caveat deserves the last word. There is a real asymmetry hiding inside the celebrations of AI-found bugs: the finder only has to be right once — and then wait for millions of site owners to act — while the attacker only has to be right once and moves first. No amount of agent-driven auditing fixes the deeper problem: decades-old software maintained by volunteers, small agencies that never budget for updates, and site owners who do not know what their own websites run on. The web runs on maintenance, and maintenance has always been the part nobody pays for. AI has not changed that. It has only made the cost of skipping it arrive faster.
Sources: Wiz, technical disclosure of CVE-2026-109618 and CVE-2026-109619 (wiz.io). TechCrunch, reporting on the PHP emergency fixes and NSA guidance, 21–22 October 2026. SecurityWeek, reporting on exploitation attempts and the disclosure leak, October 2026. Microsoft, on the ConnectWise compromise attributed to Storm-0940 and the ARTEX framework, 22 October 2026. AI News Today briefing (ai-news-online.com), 22 October 2026.
