AI-assisted security research has exposed critical weaknesses in PHP, the software behind many websites. The immediate priority is updating affected systems. The wider question is how AI changes the effort required to find—and exploit—complex bugs.
Facts checked on 21 October 2026
On 21 October 2026, PHP released security updates addressing two critical vulnerabilities: CVE-2026-109618 and CVE-2026-109619. Both can enable remote code execution without authentication. That means an attacker may be able to run code on an affected server without first signing in. [1]
The findings, discussed by Wiz under the name PHP2Shell, also offer a concrete case study in AI-assisted cybersecurity. Understanding the discovery requires looking at both the researchers’ work and the software underneath a website.
The vulnerability sits below the website
PHP is the runtime that executes an application’s server-side code. The two advisories describe failures in how that runtime manages memory and internal state when processing requests. For the first vulnerability, an out-of-memory error can interrupt an operation before its data is properly initialized. For the second, a worker can retain references to memory belonging to an earlier request after that memory has been freed. [2] [3]
A useful analogy is a shared desk between shifts: the next person should receive a clean workspace, but information from the previous shift remains and is mistaken for valid instructions. In software, this kind of confusion can have consequences beyond a simple crash.
This distinction matters for WordPress owners. Updating a theme or plugin does not replace the PHP runtime. The security response therefore needs to include the hosting environment, even when the application itself is fully updated.
How researchers used AI
Wiz reports using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol in its research. It says AI helped with discovery, exploit development and patch validation, including the difficult work of adapting to different memory layouts and configurations. These are the research team’s findings about its own workflow, rather than an independent benchmark of the models. [4]
There was also an independent discovery. According to OpenAI, Calif identified a critical flaw using GPT-5.6 Sol through the Patch the Planet initiative, while Wiz independently discovered and reported both vulnerabilities. PHP’s maintainers led the fixes and release. [5]
For background on one of the models involved, see DemystIA’s explanation of Gemini 4 Argon and its practical uses.
Why complexity may offer less protection
Our reading of this case is that the important change concerns the cost of investigation. Recognizing a suspicious code path, reproducing a failure and establishing its security impact are separate jobs. A model that helps researchers iterate through those jobs could make a previously expensive investigation practical.
That possibility cuts both ways. A defender may uncover a weakness sooner; an attacker may need less effort to adapt a technique. The useful question for a business becomes how quickly it can move from an advisory to a verified fix.
The published accounts do not provide a controlled comparison proving that one model is universally better at security research. A successful investigation combines model capabilities, tools, access to the relevant code and expert judgment. Its result should be assessed as a complete research workflow.
Which PHP releases contain the fixes?
The PHP project lists the following corrected releases for its supported branches on 21 October 2026. Website operators should obtain the applicable update through their hosting provider or software distributor. [1]
| PHP branch | Corrected release |
|---|---|
| 8.2 | 8.2.35 |
| 8.3 | 8.3.36 |
| 8.4 | 8.4.27 |
| 8.5 | 8.5.12 |
PHP says that versions 8.1 and older no longer receive regular updates from the project. Some distributors provide backported fixes, and the PHP team has exceptionally supplied minimal source backports for 7.4, 8.0 and 8.1. Those backports do not mean every hosting package for an older branch has been updated. [1]
What website owners should check
The PHP project says updating is the complete remedy; temporary measures only reduce exposure to particular attack paths. [1] A practical response should establish three things:
- Who owns the update? On managed hosting, ask the provider to confirm that the PHP service running your website includes the fixes. On a self-managed server, assign the update to the person responsible for the runtime.
- Which environment was checked? A terminal’s PHP version can differ from the version serving web requests. Verify the actual website runtime and the relevant package or provider advisory.
- What proves completion? Confirm that the updated service is running, then check essential site functions. Keep a record of the installed package and the provider’s confirmation where applicable.
These are operational checks, not an invitation to run exploit demonstrations against a production website.
The next challenge is getting fixes into use
OpenAI’s account describes a coordinated response involving maintainers, researchers and infrastructure providers before disclosure. It highlights the work of testing fixes, preparing protections and reaching the organizations that need to act. [5]
For DemystIA, that is the lasting business lesson: investing in AI-assisted discovery also requires investing in the ability to respond. An accurate inventory of software, clear responsibility for updates and a way to verify deployment give a security finding somewhere useful to go.
Sources
- PHP — Security advisory: Critical remote code execution vulnerabilities, update now, 21 October 2026.
- PHP — RCE via OOM corruption (CVE-2026-109618), 21 October 2026.
- PHP — RCE via stale global state (CVE-2026-109619), 21 October 2026.
- Wiz Research — PHP2Shell research report, 21 October 2026.
- OpenAI — Patch the Planet: global response coordination as agents uncover critical vulnerabilities, 21 October 2026.