AI Change Tracker
4

AI-assisted discovery of critical PHP vulnerabilities triggers emergency internet-wide patch and mass exploitation

2026-10-21

Security researchers at Calif, working with OpenAI's "Patch the Planet" program using GPT-5.6 Sol, and separately at Wiz (part of Google Cloud, using Gemini 4 Argon, GPT-5.6 Sol, and Wiz's own Wiz Atlas tool), independently discovered two memory-corruption vulnerabilities in the core PHP engine (CVE-2026-109618, CVE-2026-109619), with demonstrated pre-authentication remote code execution against WordPress, Drupal, and Joomla. Wiz found the first flaw around August 19, 2026 and reported it to the PHP security team August 24; the PHP project spent roughly two months on a coordinated fix. The planned October 22 disclosure was moved up to October 21 after the CVE IDs (with an incorrect CVSS score of 10.0; the actual score is 9.2) leaked onto social media and a proof-of-concept video from Calif was briefly published and copied before deletion. Cloudflare's CTO reported exploitation attempts rising from thousands in the 10 hours before public disclosure to millions within hours of the patch release, against infrastructure running PHP, which powers an estimated 70% of the web.

Significance 4: A landscape-level demonstration that multiple frontier models can independently find high-severity, previously-unknown vulnerabilities in widely-deployed software, with real two-month coordinated-fix consequences and mass exploitation after a disclosure leak across roughly 70% of the web's PHP footprint — held at 4 rather than 5 since no verdict or lab strategic position changed.

technology operational business Safety / alignment Tooling / agents

Implications · machine-drafted, not owner judgment

This blurs the line between AI as a defensive tool and AI as an attack accelerant: the same capability that let OpenAI and Wiz find and help patch the bugs also lowers the bar for less sophisticated attackers to find the next one. For the owner's enterprise-adoption lens, this is a concrete case of AI capability outrunning organizational readiness — most patch-management processes weren't designed for exploitation scaling from thousands to millions of attempts within two hours of disclosure. Expect security vendors and CISOs to start treating 'AI-discovered vuln' as a standing threat category rather than a one-off story.

Watch for
  • Whether additional AI-discovered critical vulnerabilities surface in major software within the next month
  • Independent or audited accounts from OpenAI or Google of their models' vulnerability-discovery methodology
  • Whether the PHP project or another affected ecosystem changes its coordinated-disclosure process in response to the leak

Sources

Models

← Back to feed