OpenAI has open-sourced php2shell-checker on GitHub, giving site administrators a way to check whether their PHP versions fall within the affected range of the php2shell flaws. The project went live on Oct. 21, Taipei time, under the Apache 2.0 license, and had collected more than 700 stars by the evening of Oct. 22, Taipei time.
The tool targets CVE-2026-109618 and CVE-2026-109619. According to the source material, both PHP runtime vulnerabilities carry a CVSS score of 9.2, and an attacker does not need to log in before sending crafted HTTP requests that may lead to arbitrary code execution on a server.
Codex skills and a Node.js option
OpenAI was part of the group that disclosed the php2shell vulnerabilities. With this release, it has moved from disclosure into tooling for defenders. The project documentation says users can ask Codex to install two skills, phpcheck-url and phpcheck-server.
phpcheck-url only needs a website address, which makes it suitable for shared hosting users without server access. phpcheck-server runs php -v inside the server environment to inspect the version that is actually in use.
Users who do not want to rely on an AI agent can download the project and run phpcheck.js with Node.js. The checker returns three possible results:
- OK
- VULNERABLE
- UNKNOWN
Each result maps to a different exit code, and the tool also supports JSON output, making it easier for operations teams to plug it into automated batch scans.
No exploit traffic, but also no proof of remediation
The checker is deliberately designed not to send exploit code. In URL mode, it only sends a HEAD request, reads the Server and X-Powered-By headers, does not follow redirects, and does not parse page content. OpenAI says the load on a checked site is close to that of a normal visitor.
That design choice reduces risk during scanning, but it also sets clear limits on what the tool can confirm. In its documentation, OpenAI says an OK result is only a version-number comparison. It assumes that later stable patch releases in each branch include the fix, but the checker does not independently verify whether the patch is actually present and working.
OpenAI says UNKNOWN does not mean patched
OpenAI also notes that many websites hide or rewrite response headers, while Linux distributions often patch vulnerabilities through backporting without changing the upstream version number. That means UNKNOWN results may not be rare.
The documentation highlights this point with an explicit warning: 「UNKNOWN 不代表已修補」. In practice, the checker is closer to a first-pass filter than a clean bill of health. For administrators with server access, the source material says phpcheck-server results and official PHP security advisories should carry more weight.
CISA calls for upgrades, Cloudflare says it saw attack traffic
The U.S. Cybersecurity and Infrastructure Security Agency, or CISA, said in an Oct. 21 alert that users should immediately upgrade to PHP 8.5.12, 8.4.27, 8.3.36, or 8.2.35. Systems still running unsupported versions below 8.1 should be taken offline first, according to the alert cited in the source material.
Cloudflare said in a company blog post that the vulnerabilities were jointly disclosed by Calif.io, Wiz, OpenAI, and Quarkslab, and that information had leaked before the planned disclosure date. Cloudflare also said it enabled protections by default for both free and paid WAF users starting at 11:02 a.m. Taipei time on Oct. 21, and that it had already observed attack traffic.
Security firm watchTowr likewise said it had seen exploitation attempts. CISA, however, said when it issued its alert that it had not yet observed active exploitation. The source material notes that the statements were made at different times, and the actual scale of impact could not be independently verified.
Why crypto companies may care
Cloudflare says nearly 70% of websites use PHP, and platforms such as WordPress and Magento are built on it. If exchange websites, project blogs, or some front-end pages rely on that stack, a compromise could turn them into a vehicle for phishing or wallet-address tampering.
The source material also draws a contrast with incidents such as Log4Shell, where the response was led mainly by cybersecurity firms. In this case, an AI lab is present on both the disclosure side and the tooling side, showing how frontier models are being positioned as part of defensive infrastructure.
php2shell-checker does not patch anything on its own. The actual fix is still to upgrade PHP. What OpenAI has done is package the checking process into skills that an AI agent can run directly, lowering the barrier for smaller site operators. For crypto websites and project teams that are frequent targets, the sequence described in the source material is straightforward: inventory PHP versions first, then combine that with WAF protection and official patches.


