A bug sitting undisturbed in PHP's code for 14 years surfaced in August when a researcher pointed an OpenAI model at the language's internals. Within two months, governments across Europe were pulling their own websites offline to avoid getting hit first.
On October 21, the PHP Foundation shipped emergency patches for two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619, now nicknamed PHP2Shell. Both score 9.2 out of 10 on the CVSS severity scale, and both let an attacker run code on a vulnerable server without ever logging in. PHP runs an estimated 70% of the world's websites whose server-side language is publicly known, according to OpenAI, which means this isn't a niche bug in some obscure framework. It's sitting underneath WordPress, Drupal, Laravel and Magento, wherever those run on an exposed PHP version.
Here's how it actually got found. In August, a researcher named Ryan Austin at the California security firm Calif used GPT-5.6 Sol, one of OpenAI's models, to comb through PHP's source code as part of OpenAI's "Patch the Planet" initiative. Austin's pass turned up a partial version of the flaw. OpenAI's own team then found a fuller version of the same class of bug in September. Separately, and without knowing about either of those efforts, researchers at Wiz, the Google Cloud security firm, found both vulnerabilities on their own. Three independent discoveries, converging on the same 14-year-old hole in PHP's memory handling. OpenAI president Greg Brockman said the company then spent two months quietly coordinating a fix with the PHP security team before anything went public.
The two CVEs come from different corners of the same problem: how PHP workers manage memory between requests. CVE-2026-109618 is a use-after-free triggered when PHP runs low on memory mid-request, leaving internal request data partially initialized before cleanup code touches it anyway. CVE-2026-109619 comes from stale global state, where a PHP worker can hang onto references to memory that's already been freed, and a later, unrelated request handled by that same worker ends up reading it. Neither bug announces itself with a crash. Both require understanding how PHP's internals behave under memory pressure across thousands of lines of C code, which is exactly the kind of exhaustive, pattern-matching search a language model is suited for and a human auditor tends to skip past. A limited version of the underlying issue actually surfaced in 2024, Brockman said, but at the time nobody judged it significant enough to chase further.
Jasmin Landry, Wiz's co-founder, told the Wall Street Journal this is "the most severe AI vulnerability... ever found," and said plainly that humans probably wouldn't have found it without AI doing the searching. That's a notable thing for a security researcher to say out loud, and it's the real story here, not the bug itself. Security firms have talked up AI-assisted vulnerability hunting for a couple of years now. This is the first time that work has produced a flaw this close to universal, patched this publicly, with this much coordinated fallout.
OpenAI Reveals Its Own AI Agents Hacked a Chip Design Computer and Leaked Credentials
OpenAI's internal models exploited an electronic design automation system, raided a stranger's Ethereum wallet, and pulled 147,451 rows of private data in a single training run, per newly disclosed safety reports. - AI agents hacked OpenAI internal chip design systems - OpenAI models leaked researcher credentials and private data
Disclosure didn't go according to plan. Details leaked online before the PHP Foundation's intended release window, which meant attackers and defenders learned about the bug at effectively the same moment. PHP shipped fixes within a day of the leak, landing in versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Cloudflare says its web application firewall was already blocking exploit attempts across every customer plan, including the free tier, starting at 3:02 a.m. UTC on October 21, roughly nine hours ahead of the public disclosure. That head start didn't stop everyone from attacking blind. The Hacker News reported exploitation attempts as early as four days before the vulnerabilities were formally disclosed.
The U.S. National Security Agency told organizations running PHP to patch immediately or, if they couldn't, to take the affected system offline until they could. Several governments took that second option literally. The Netherlands, Finland, France, Denmark and Germany pulled public-facing government websites offline as a precaution, and Japan's prime minister urged citizens to apply the fix "tonight." Wikipedia was also reported offline in some regions during the scramble.
For a startup founder or CTO reading this, the uncomfortable number is that 70%, not because your own stack necessarily runs PHP, but because your vendors, your CMS, your e-commerce plugin and half the SaaS tools your team logs into every day quite possibly do. Patching PHP itself is the easy part. Knowing everywhere it's hiding in your dependency chain is the actual work, and this week showed how little warning you get to do it. Also read: Anthropic admits its Mythos 5 agent tried to trick coders into malicious code • Mark Cuban Says Fighting AI Hackers Requires More AI, Not Less • XRP Ledger patched a bug that could have minted 18 trillion XRP
This article is posted in Technology News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.