WordPress just went AI

About 41% of all websites worldwide run on WordPress.¹ So when WP builds AI into the core of the platform, a large part of the web gets access to AI in one update.

What does that actually change, how fast, and what does the rollout tell us about how AI enters the tools companies already use?

Before version 7.0, adding AI to a site meant patchwork. Each plugin built its own connection to an AI model, with its own keys and its own rules, so a site with three AI plugins had three separate setups and three places to control them. Version 7.0 moved that connection into the core.² The site owner now links an account at a company that supplies AI models, such as OpenAI, Anthropic or Google, only once, and every plugin built for the new system uses that link. Alongside the release, an official adapter gave AI agents a way to read and edit the site’s content directly.³

The core itself contains two pieces:

• the shared connection to AI models,
• and a register of site functions that AI agents can find and use.
• Ready-made features, such as generating titles, excerpts or image descriptions, come in a separate, optional plugin, and everything else is left to plugin developers.³

Since new AI features now depend on those developers, the change will move at their pace. The first steps are already under way:

• In August, version 7.1 let plugins add their own connections to other AI services.⁴
• Alongside the next version, due in early December 2026, the team is working on agents that can write content as well as read it, and on giving each agent its own identity and permissions, separate from human accounts.⁵

For now, that work lives in a separate plugin.⁵ Each release lets AI do a little more inside a site, so the full impact on how sites are built and run will take years.

While the AI connection was being added to the core, the platform’s older code came under attack:

• In July 2026, the core team had to patch two flaws that, chained together, let anyone take over a site without logging in, and attackers used them widely.⁶
• In August 2026 came twelve more fixes, some for weaknesses present since 2016.^{7 8} The holes sat in code older than the AI features, some of it by nearly a decade, so AI’s part in this story was finding them.

Researchers turned the July chain into a working attack by setting an AI model on the platform’s code. It took them ten hours and about $25.⁹ The same kind of model that the platform now invites inside is reading its code from outside, and weaknesses that stayed hidden for years surface in hours.

Anyone connecting AI to an old system has to assume the old parts will be tested harder than ever.

Companies adding AI to their own older software face the same choice the core team made.

A system that has grown for twenty years carries rules, exceptions and history that nobody fully remembers. Rewriting it is too risky and too expensive, so AI is attached to it instead.

It connects to what exists and steers it from the outside, while the foundation keeps running.¹⁰ The platform, at twenty-three years old, is a very public case of what ERPs, CRMs and warehouse systems are going through quietly.

Attaching AI to working systems changes the question from “can we add AI here” to “does it pay off here”. The WP core team has turned that into a rule. During the 7.1 cycle, it decided that an AI feature enters the core only once it proves real adoption and practical value, and until then it stays in a plugin.⁵

The result is that AI will be used where it earns its place, and AI agents will act within the permissions people set for each of them.