AI Found a Decade-Old Hole in PHP. Now 70% of the Web Is Racing to Patch.

Two flaws in PHP, the language behind roughly 70 percent of websites, sat unnoticed for a decade until frontier AI found them. On Wednesday governments pulled sites offline while the world applied the fix.

On Wednesday morning, the Finnish government took its public web services offline. So did the Dutch public health institute and the lower house of the Dutch parliament, along with parts of Wikipedia. None of them had been breached. They were pulled down because of two bugs in PHP — the scripting language that quietly runs a huge share of the web — and a patch that landed that same morning. Edinburgh council's website and the online homes of several BBC programmes also went dark, though neither has confirmed a PHP link.

The flaws are indexed as CVE-2026-109618 and CVE-2026-109619, and both allow unauthenticated remote code execution, with severity reaching CVSS 9.2. They are not new. Wiz Research says every PHP release of the past decade carries at least one of them, going back to version 5.4. Exploiting them required chaining memory-corruption bugs across requests, in ways that depended on how PHP allocates and reuses memory — "an exceptionally complex chain of steps, at the edge of what human researchers could practically achieve," the firm wrote. It packaged the result into a single exploit it calls php2shell, which it demonstrated against default container images of WordPress, Drupal and Joomla.

What changed was who found them. Calif, a security firm working inside OpenAI's Patch the Planet program, identified one flaw using GPT-5.6 Sol. Wiz independently discovered and reported both, using its own Atlas system alongside Gemini 4 Argon and GPT-5.6 Sol. An OpenAI researcher hit the same issue at roughly the same time. The bugs had survived more than a decade of human review; the AI-assisted hunt took a few weeks.

The exposure is enormous by any measure. W3Techs puts PHP on 69 percent of websites, and Censys counted nearly 55 million web properties running it as of Wednesday. Wiz's headline says 70 percent of the internet is affected, though that figure describes exposure rather than confirmed compromise — the BBC said it could not verify it, and researchers cautioned that practical exploitability varies by configuration. The vulnerable code, however, sits beneath content management systems, hosting panels and third-party products that organisations may not even know they run.

The disclosure window did not hold. The PHP Foundation had planned to publish later this month alongside a complete fix; instead, details leaked on social media on Tuesday and forced the announcement forward. NBC News, citing the foundation, reported that Chainguard posted details of the patch six hours before its release, breaking an embargo and handing anyone watching a head start on reverse-engineering the fix. Cloudflare's chief technology officer said he was "seeing early attacks emerging and are updating mitigations as new vectors show up." Patchstack said it had received reports of active exploitation; the US Cybersecurity and Infrastructure Security Agency said it had not observed exploitation as of its advisory.

Governments moved quickly in the other direction. The National Security Agency told organisations to take systems offline if they could not patch immediately, warning the flaws "often provide a direct path to the server's file system or database." Britain's National Cyber Security Centre and CERT-EU issued similar guidance, Japan's prime minister urged the public to check whether the sites they rely on run PHP, and the White House said federal agencies were working with leading AI labs to speed a fix.

The patched releases are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, with minimal source backports for 7.4, 8.0 and 8.1; anything older has to be upgraded. The foundation's own advice is blunt: "There are no complete mitigations, updating is the only safe choice." The harder problem is inventory. As Sophos's Don Smith put it, PHP "can sit beneath public websites, content management systems, hosting platforms and third-party products, sometimes without even being obvious to the organisation using it."

OpenAI used the moment to argue for a new doctrine. It published a five-stage framework for disclosing critical vulnerabilities in widely used software, said it had coordinated with the White House's Super Intelligence Force, the Office of the National Cyber Director and CISA, and committed $1 billion in subsidised access to its security models for defenders with limited resources. That the company whose model found the bug is also drafting the rules for disclosing such bugs is worth watching. The framework's own logic concedes the tension: the same capabilities that let defenders find flaws first let attackers build exploits faster.

The bugs were present through every audit, every penetration test and every rewrite of the last ten years. AI found them in weeks, and the patch took a morning. The patching will take much longer. Alan Woodward of Surrey University expects "a long tail of websites who may not be maintained as well" — and with it, in his words, "weeks and months of ongoing attacks."

Sources: BBC, The Guardian, NBC News, Wiz Research, PHP.net security advisory, NCSC, CERT-EU