Two PHP vulnerabilities discovered by AI expose a large part of the Web

Two critical vulnerabilities in PHP were identified with AI assistance and already exploited before disclosure. Patches exist, but part of the Web could remain exposed for a long time.

Two critical vulnerabilities in PHP, the language powering a large portion of websites, were discovered thanks to artificial intelligence and were already exploited by hackers before their public disclosure. Published under the references CVE-2026-109618 and CVE-2026-109619, they allow an unauthenticated attacker to execute arbitrary code on a server. The issue is immediate for administrators but also long-lasting for sites that cannot migrate quickly.

PHP au cœur du problème, pas à la marge

Researchers from OpenAI, Wiz, and Calif used GPT-5.6 Sol to analyze PHP code, a widely used server-side language. Their discovery affects the core functioning of the engine: according to the PHP Foundation, the centrality of these components makes the fix delicate, as compatibility, performance, and the extension ecosystem must be preserved. As a result, the vulnerability does not target a peripheral module but a layer that concerns almost all requests.

The scenario is classic and frustrating at the same time: a vulnerability becomes public, then attackers quickly seek to exploit it while technical teams apply updates. Here, the window has already started to close even before the announcement. Amazon reportedly observed attacks as early as October 17, before the publication on October 21. Hackers can then dissect the patches to build new tools. The timeline is somewhat harsh but at least very clear.

Des correctifs existent, mais pas pour tout le monde

The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35, depending on the branch used. PHP.net also published backporting code for PHP 7.4, 8.0, and 8.1, showing the scale of the problem for older installations. The vulnerabilities have reportedly been present since at least PHP 5.4 [to be verified], and many sites continue to run outdated versions, sometimes because an old third-party script blocks migration.

This technical inertia weighs heavily. Many sites depend on hosting providers, do not have control over the server, or no longer have the skills to audit their scripts. Others use CMSs like WordPress, Drupal, or Joomla, as well as Roundcube, Nextcloud, and phpMyAdmin, all potentially inheriting this risk if the underlying infrastructure is not updated. According to the source text, about 70% of the Web could potentially be affected [to be verified], giving an idea of the scope without reducing the alert to a simple routine patch.

Quand la mise à jour devient une course contre l’usage

The real issue is not only the vulnerability but the actual deployment speed. Some sites can be fixed quickly; others will remain exposed for weeks, months, or even years if updates break critical dependencies or if the administrator simply no longer has access to the server. In this context, AI has not only helped find the flaw: it also accelerates the attack-patch-counterattack cycle.

For organizations, the message is simple: check the PHP version, apply available patches, and document systems still at end-of-life. For the rest of the Web, this episode reminds us that old software likes bad surprises. And it keeps them for a long time.

Key points

  • Two critical CVEs: 2026-109618 and 2026-109619.
  • OpenAI, Wiz, and Calif used GPT-5.6 Sol.
  • Amazon observed attacks as early as October 17.
  • PHP.net published patches on October 21.
  • PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35 are affected.

By the numbers

  • 2 vulnerabilities — discovered in PHP, according to the source published on October 22, 2026.
  • 4 fixed versions — 8.5.12, 8.4.27, 8.3.36, and 8.2.35.
  • 70% of the Web — potential exposure advanced by the source [to be verified].
  • October 17 — first attack observed by Amazon before the announcement.

To read