Technology

PHP2Shell: Critical PHP Flaws Patched After AI-Found Bugs

Martin HollowayPublished 3d ago4 min readBased on 10 sources
Reading level
PHP2Shell: Critical PHP Flaws Patched After AI-Found Bugs
Photo by Arnold Francisca on Unsplash

The PHP development team released emergency security updates on 21 October 2026 to fix two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619, which could allow unauthenticated remote code execution. Help Net Security reported the emergency releases, and PHP had scheduled its critical security release for 14:00 UTC that day.

The flaws were disclosed by Wiz Research. Wiz published its write-up on 20 October under the title "PHP2Shell: AI uncovers critical RCE vulnerabilities," describing core PHP engine flaws that enable remote code execution and Heartbleed-style memory disclosure. According to Wiz, the first bug was found on 19 August 2026 using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol. The remote code execution vulnerability was discovered entirely by AI agents, a detail Wiz's research team walked through in a follow-up discussion published the same day.

The exposure is large. PHP is the open-source language behind roughly 70% of websites, so a core engine flaw reaches a wide range of hosts, content management systems and custom applications, many of them maintained by small teams with limited patch cycles. Some organizations took their sites offline until the vulnerabilities could be patched, according to TechCrunch's coverage of the disclosure. Reuters reported on 22 October that hacking attempts surged as companies and governments rushed to patch the two flaws, and that Cloudflare had seen millions of exploit attempts following the PHP disclosure.

Wiz Research has been monitoring exploitation attempts since disclosure. In a follow-up analysis published on 27 October, titled "PHP2Shell: Exploitation Attempts and Remediation Trends," the company reported that none of the attempts have succeeded so far.

Google Cloud also published a response to the two vulnerabilities on 20 October. Wiz is owned by Google, and Wiz's CTO of Exposure Management and AppSec, Tomer Schwartz, said the company worked alongside OpenAI, the research group Calif, and the PHP team to fix the bugs. Schwartz was interviewed by Rebecca Bellan for an episode of TechCrunch's Equity podcast on PHP2Shell and what the episode called the "coming vulnpocalypse." Wiz is also hosting a 20-minute technical threat briefing with its research team for organizations assessing their exposure.

The coordination itself is the part of this story that matters most to operators. A disclosure that involves a core language runtime, multiple AI labs and a commercial security vendor has to move through embargo, patch development and downstream packaging before the details become public, and the fixes had to reach hosting providers and application owners quickly enough to matter. Wiz's account places the discovery in August and the public disclosure in late October, and the emergency patch window that followed was short. Whatever else the episode shows, the patch pipeline (vendor release, distribution to hosts, application-level testing and restart) is where the time went.

In my view, the more durable question is what this does to the economics of vulnerability management. If AI agents can surface critical remote code execution bugs in a language this widely deployed, the bottleneck moves from finding flaws to deploying fixes across a long tail of systems that nobody has inventoried. Organizations that know exactly which PHP versions they run, and can roll out a patched build within hours, will absorb this with little drama. Those that cannot will face the same problem again the next time the disclosure cycle turns over.

The broader context is also encouraging, with a caveat. The discovery route described by Wiz is the same tooling that the industry has been debating for two years, and here it found bugs in shared infrastructure that the community could then fix in coordination with the language maintainers. Monitoring has so far found no successful exploitation, and the patches are in the hands of the people who run the sites. The open question is whether the pace of AI-assisted discovery will outrun the slow, unglamorous work of patching, and that will be settled by operators, not by the disclosure itself.