Back
aiproduct

AI finds vulnerabilities faster than they get fixed: PHP2Shell lessons

Critical PHP flaws were found with AI. Why the window between discovery and attack is shrinking, and what that changes in how business systems are maintained.

Published: 2026-10-23

On October 21 PHP shipped emergency updates for two critical vulnerabilities known together as PHP2Shell. Both were found with the help of AI, and exploitation attempts began the same day. The short answer to "what does this change": updating business systems stops being a project every few years and becomes routine work with agreed deadlines.

The PHP2Shell vulnerabilities

The PHP project released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, fixing CVE-2026-109618 and CVE-2026-109619. Both are rated 9.2 out of 10 and allow unauthenticated remote code execution. The flaw is in PHP itself, not in the site's code, so anything running on affected versions is potentially exposed — WordPress, Drupal, Russia's popular 1C-Bitrix platform, custom customer portals. Wiz demonstrated working attack chains against current WordPress, Drupal and Joomla images. There are no complete mitigations; the project says plainly that updating is the only safe choice.

Branches 8.1 and older are no longer supported and got no official release. According to Shodan data cited by CybelAngel, of roughly 1.5 million internet-facing servers that report their PHP version, 59% run a branch with no official fix.

AI's role in the discovery

The first issue was reported on August 11, independently, by researchers at Calif and by OpenAI. Calif takes part in OpenAI's Patch the Planet program, where security engineers work with models to find and fix bugs in open-source software. Wiz later found additional attack paths using its own tooling and Gemini and GPT models. Wiz says the models helped at every stage: discovery, building the exploit chain and validating the patches.

The most telling part is the history of the original bug. According to Calif, the weakness in out-of-memory handling was found back in 2020 and again in a 2024 audit, but stayed unfixed because nobody proved it was exploitable. Current models turned it into working remote code execution. Wiz describes the chain of steps as being "at the edge of what human researchers could practically achieve."

This is not a one-off. The PHP Foundation says AI is helping surface more vulnerabilities, with 15 new PHP CVEs published since July 1. In the foundation's words, tools now find bugs faster than ever, for defenders and attackers alike.

The patch window is shrinking

Disclosure was planned for October 22, but details leaked on the 20th and PHP released the fixes a day early. The same day watchTowr, Patchstack and Cloudflare saw the first exploitation attempts; Amazon, as reported by SecurityWeek, saw attempts four days before the updates. No successful intrusion had been confirmed at the time of writing, but watchTowr does not expect that to last.

The response was unusually strong: the US National Security Agency advised taking sites offline if they could not be updated immediately, and Germany and France took some official sites offline while patching.

For businesses this leads to three conclusions:

  • The time between a fix being published and an attack is now measured in hours. "We'll update next quarter" no longer works for critical vulnerabilities.
  • Old, known issues become dangerous. A bug considered theoretical for years can become a real threat once a model takes it apart.
  • Unsupported versions are a risk with no fix. If your platform no longer receives patches, there will be nothing to install when the next vulnerability lands.

The risk for systems that are not updated

Many small and mid-sized business systems run on "if it works, don't touch it": a site on a CMS, a customer portal written five years ago, an integration with accounting that hasn't been updated since launch. They don't break on their own — they get broken when a platform vulnerability appears and they can't be updated quickly.

The defenses are well known and don't require AI: a supported platform version, someone responsible for updates, and an agreed timeframe for installing them. The answer to "who will update PHP tonight" should be known in advance.

AI helps defenders too: the same models can review your own code and dependencies and validate fixes. But deciding what to fix and how to roll it out stays with people — in Patch the Planet and at Wiz the model finds, and a specialist verifies. We build AI agents on the same principle.

Checking your own systems

  1. Check versions. Ask your hosting provider or contractor which PHP version your site, portals and CRM run on. Anything below 8.2 calls for a separate decision, not waiting.
  2. Put deadlines into your support contract. How many hours or days to install a critical update, who tracks advisories, how you will be notified.
  3. Keep a list of systems and their platforms. Site, portal, integrations, on-premise CRM — what each runs on and until when it is supported. PHP 8.2, for example, stops receiving fixes on December 31, 2026.
  4. Decide what to do with systems on old versions. Update, gradually move to a supported platform, or take them off the public internet.
  5. Use AI to review your own code, with a specialist going through the findings.

SLA-based support with fixes in agreed timeframes and monitoring is part of our work on ERP and business systems. If a system came from another contractor, we start with an audit of platforms and versions.

Sources