AI Save the Internet and You

AI helped researchers expose critical weaknesses in PHP, software behind many everyday websites. Protection came from a team: researchers established the danger, maintainers built fixes, and website operators had to install them.

What does “save the internet” mean here?

It means helping uncover and repair serious weaknesses in shared software before they can cause further harm. It does not mean that AI alone fixed every website, that every attack was prevented, or that the internet is now immune to intrusion. The patch effort continues wherever affected software remains in use.

What is PHP2Shell?

PHP’s 21 October security advisory identifies two critical issues in the engine’s memory and request handling. Under affected conditions, they can allow an attacker to execute code on the server without logging in. Wiz also demonstrated sensitive process-memory disclosure, potentially exposing credentials such as passwords and access tokens. These are server-side flaws, not a browser feature that visitors can repair themselves.

What did AI contribute?

OpenAI’s 21 October account says Calif identified a critical flaw using its models through the Patch the Planet initiative. Wiz’s independent research describes using AI to discover, test and adapt demonstrations across different PHP applications. These accounts support AI-assisted research, not a claim that an autonomous system repaired the web without people.

The PHP Foundation’s account adds important history: concerns about the out-of-memory root cause had surfaced in 2024. Demonstrating practical impact was difficult, and choosing a compatible fix involved substantial engineering trade-offs. The 2026 reports made the danger clearer and brought the repair to the top of the team’s priorities.

Why could this affect ordinary users?

PHP helps servers generate web pages and operate applications. Visitors may use a PHP-backed service without knowing which language it runs. A server compromise or memory leak can therefore put the service’s data at risk, even if users have done nothing wrong. A possible exposure is not evidence that a specific person’s credentials were stolen.

The widely repeated “70% of the web” figure also needs its denominator. W3Techs’ 22 October snapshot reports PHP on 69.3% of websites whose server-side programming language it can identify. That is a technology-usage estimate. It does not measure the fraction unpatched, exploitable in a particular configuration, or successfully attacked.

What happened before public disclosure?

Amazon Threat Intelligence’s 21 October report traces an attempt to 17 October, using telemetry from MadPot, its decoy-system network. It also describes further attempts before and after disclosure. Amazon explicitly says it is unclear whether the observed attempts would have succeeded against real-world systems.

Those observations support investigating activity that predates the advisory. They do not identify a worldwide number of victims or prove that an attacker used AI. AI-assisted defensive research and observed malicious attempts are separate facts.

Which releases contain the fixes?

Source: PHP’s official advisory and release notices. Distribution and long-term-support vendors may backport fixes while retaining older-looking version numbers. PHP 8.1 and older no longer receive normal upstream updates; ask the vendor about a supported fix or migrate to a supported branch. A version banner alone is not enough to assess a vendor-patched installation.

What should operators and visitors do?

The UK NCSC’s guidance advises operators to identify PHP installations, apply updates, monitor vendor advisories and investigate possible compromise. Interim mitigations have limits; follow the official guidance when deciding whether an exposed service should remain online. Verify that the updated software is actually running. Installing a patch does not, by itself, remove an earlier intrusion.

For visitors, the server update belongs to the service operator or hosting provider. Follow the service’s official notices if it reports an incident. This explainer has not scanned your browser, device or accounts and makes no claim that they were breached.

Primary sources and dates

1. PHP official advisory and release notices, 21 October 2026: CVEs, affected server deployments, fixed releases and vendor backports.
2. PHP Foundation: today’s security release, 21 October 2026: earlier root-cause concerns, human engineering and disclosure history.
3. OpenAI: coordinated disclosure and Patch the Planet, 21 October 2026: Calif’s AI-assisted discovery and response coordination.
4. Wiz: PHP2Shell research, 21 October 2026: independent findings, AI-assisted testing and demonstrated impacts.
5. Amazon Threat Intelligence: MadPot observations, 21 October 2026; earliest reported attempt 17 October. Real-world success of observed attempts remained unclear.
6. UK NCSC: critical PHP security issue, published 21 October 2026, checked 23 October; operational response guidance.
7. W3Techs PHP usage, snapshot dated 22 October 2026. Denominator: surveyed sites with an identified server-side language.

Research checked 23 October 2026. This is a dated snapshot, not a live incident dashboard. Later guidance may change. Illustrations show concepts, not actual network activity or measured attack volumes.