Wiz used three AI models to find two use-after-free bugs in PHP

Wiz Research published a disclosure for two remote code execution vulnerabilities in PHP's unserializer, CVE-2026-109618 and CVE-2026-109619, both scored 9.2. The authors are Sagi Tzadik, Shir Tamari, Nir Ohfeld and Tomer Schwartz. PHP's security team shipped patches for the bugs the same day, which are the subject of a separate release note. The notable material in the Wiz write-up is the method of discovery.

The three models named

Wiz says three AI models took a share of the work. Gemini 4 Argon from Google DeepMind, GPT-5.6 Sol from OpenAI, and Wiz Atlas, the company's own internal tool. The authors state that Gemini 4 Argon made "an invaluable contribution to finding the vulnerability, exploit development, attack detection, and validation of the patch."

Both bugs are use-after-free flaws. CVE-2026-109618 sits in the upload-progress unserializer and allows remote memory disclosure and code execution. CVE-2026-109619 sits in the fatal-error backtrace, triggering OPcache type confusion and code execution. Every PHP version from 5.4 forward is affected, which Wiz says covers around 70 percent of websites, including WordPress, Drupal, Joomla and Wikipedia's MediaWiki.

The timeline

The disclosure records four dates. Wiz first found the vulnerability on 19 August 2026. It reported the flaw to the PHP security team on 24 August. PHP acknowledged the issue and started the patch on 21 September. The patch and the public disclosure both shipped on 21 October. Two months from discovery to a public patch.

What the authors say that is quotable

The piece ends on a sentence worth carrying over verbatim: "As the time and expertise required for this work decrease, complexity may no longer provide a meaningful barrier, changing how we define what is vulnerable in practice."

That is the argument of the whole write-up. The specific use-after-free paths in PHP's unserializer have been in the engine for 22 years. The code review tools that found them are new, which is why the bugs turned up now.

What a working team should take from this

Three practical steps. For a security team at any company that writes C or C++ in production, the Wiz piece is a worked example of the current state of model-assisted bug hunting. Pointing a frontier model at an existing attack surface is a research step that costs some tokens and some human review time, and may surface a flaw that two decades of manual review missed. The method is public.

For a product team shipping a PHP application, the response is unchanged. Patch to PHP 8.5.12 or the equivalent on the supported branch, confirm Opcache is handled, and read the Wiz technical appendix before writing any detection rule.

For a team building AI security tools, the useful detail is that the work was split across three distinct models. The paper gives the authors' own view of what each contributed, which is a short checklist for anyone thinking of running one model instead.

Source

Wiz: PHP2Shell, how AI uncovered critical RCE flaws in PHP, by Sagi Tzadik, Shir Tamari, Nir Ohfeld and Tomer Schwartz.