Researchers from Wiz Research, OpenAI Codex Security, and Calif, with the participation of AI models Gemini 4 and GPT-5.6-Sol, discovered two critical memory vulnerabilities in the PHP core as part of the Patch the Planet initiative. These vulnerabilities allow an unauthenticated attacker to execute remote code or read process memory. The issues were assigned CVE-2026-109618 and CVE-2026-109619 and received a CVSS score of 9.2. They have existed since PHP 5.4, affecting approximately 70% of websites, including popular CMS platforms like WordPress, Drupal, and Joomla. The PHP team released emergency patches on October 21, 2026, and OpenAI published a free diagnostic tool, php2shell-checker, on GitHub to check PHP versions on servers.

image
image

What happened

Teams from Wiz Research, OpenAI Codex Security, and Calif identified two memory-class vulnerabilities in the PHP core as part of the Patch the Planet initiative and reported them to the language team through a coordinated disclosure scheme. According to the PHP2Shell project, both issues allow an unauthenticated attacker to remotely execute arbitrary code or read process memory, making any public service running an older version particularly dangerous. The PHP response was urgent: on October 21, 2026, four supported branches were closed simultaneously with releases 8.5.12, 8.4.27, 8.3.36, and 8.2.35. Such a simultaneous release across all lines indicates an emergency operational cycle rather than routine maintenance. Along with the patches, OpenAI published the php2shell-checker diagnostic tool on GitHub, which checks PHP versions across server fleets and helps identify machines with unpatched builds.

Context

PHP remains the most widely used server-side language: according to the source, it powers approximately 70% of websites, including mass-market CMS platforms like WordPress, Drupal, and Joomla. Therefore, core defects simultaneously affect hosting providers, corporate portals, and government services. The discovered errors lived in the C code of the core since PHP 5.4, meaning for more than a decade, and slipped past classic fuzzing and regular manual audits. Meanwhile, the description of a "memory error in C-code core" is plausible for a strong LLM audit, as this is a familiar class of patterns for models. The disclosure was handled through the Patch the Planet initiative, which includes Wiz, Calif, OpenAI, CISA, and the White House. OpenAI proposes establishing such a coordinated release—from report to ready patches—as a standard protocol for future AI discoveries in open-source.

Why this matters for the industry

For the industry, this is the first major precedent where AI models, in conjunction with expert teams, found a defect in the core of the most widely used server-side language that had been invisible for over a decade. The Patch the Planet disclosure scheme, involving Wiz, Calif, OpenAI, CISA, and the White House, may become a standard protocol for future AI findings in open-source, as assessed by the source. With scaling, a wave of agent-based audits of mature codebases is expected. For security teams and vendors, a ready-made product direction opens up: integrating php2shell-checker checks into control panels, CI/CD, and client onboarding, adding UX for tracking "AI findings in my stack." The wave of mass PHP checking is already driving demand for version audits. The strategic framework is dual: if the methodology is confirmed, reasoning models will become a standard stage in security audits before release and in legacy code retrospectives. However, the same toolset is available to attackers, and the asymmetry of "found earlier—patched earlier" becomes a competitive task, potentially leading to the emergence of an "AI security auditor" as a standard layer in hosting and CI/CD delivery.

Why this matters for users

If you administer servers or develop in PHP, operate in emergency incident mode: update PHP to 8.5.12, 8.4.27, 8.3.36, or 8.2.35 depending on the branch used, and restart all workers so the patch applies to already running processes. Then run the free php2shell-checker from the openai repository on GitHub across your entire machine fleet to catch servers remaining on older versions. Due to the unauthenticated nature of the attack and the public availability of the CVE with a CVSS of 9.2, the exploit window is open right now until the long tail of unpatched servers updates. Website owners using WordPress, Drupal, or Joomla only need to check with their host which PHP version is connected to their plan and request a switch to a patched branch.

What is currently unknown / limitations

The discovery methodology has not been disclosed: it is unknown what exactly the models did—whether they read code, worked in an agent pipeline, or worked in conjunction with fuzzing, what the false positive rate is, and at what stage humans from Wiz Research and Calif joined the search. This is a single case, not a proven rule: conclusions that AI vulnerability search is "sharply becoming cheaper and commoditized" and that the advantage of manual expertise in security labs is "blurring" do not follow from this, as expert teams clearly participated in the finding. The key check ahead is independent replications on other codebases: only they will show whether PHP was first in line for a systematic trend or a random stroke of luck. Finally, the figure of "approximately 70% of websites" is an estimate by the source, not the result of an independent audit of the server fleet, and the timeframe for the appearance of working exploits after the patch publication has not been publicly stated, although a short window is likely due to the public CVE and unauthenticated vector.

Sources

Author

Look at AI, editorial team