WordPress 7.0 Beta 1 is four weeks out, and the AI work is finally moving faster than the talk about it. I have watched enough core releases turn into a last-minute scramble to stay a little suspicious, but the AI Contributor group is coordinating better than I expected. This week’s meeting was mostly about the WordPress 7.0 AI features and the tooling needed to make them safe on a production site.
How the WP AI Client gets into core
The merge proposal for the WP AI Client is close to final. The plan is to bring in PHP AI Client v0.4.0 the same way the Requests library was added to core. Anyone who spent time fighting the old WP_Http code will read that as good news. It points to a real library behind a clear interface instead of another scattering of global functions.
A PR still in progress will add a filter controlling who can run AI prompts at all. Site owners need that, both for the bill and for the obvious security reasons. I made the same point when I looked at the WordPress PHP AI Client SDK: once a site has more than a couple of editors, you have to be able to say who sends what, and in which context.
The Abilities API and who gets permission
The Abilities API is meant to ship its client-side JS half in 7.0, which would let AI reach settings, users and post management. That last part is where I get nervous. Handing an LLM delete_user or update_option is asking for a very bad afternoon. The team is looking at a filtering system that limits access by user and context, which is the right instinct.
/**
* bbioon_ai_user_access_filter
* Filter to control user access to running AI prompts.
* This reflects the forthcoming PR mentioned in the weekly summary.
*/
add_filter( 'bbioon_ai_user_can_run_prompts', function( $can_run, $user_id, $context ) {
// Specifically limit AI prompts to editors in the post-editor context.
if ( 'post-editor' === $context && ! user_can( $user_id, 'edit_others_posts' ) ) {
return false;
}
return $can_run;
}, 10, 3 );WP Bench and testing the models
WP Bench is the update I keep coming back to. It is a Python tool that scores models against WordPress-specific tasks. Checking output by hand stopped scaling a while ago. Running WP Bench, or testing locally through Ollama, is how you find out whether WordPress 7.0 AI features hold up against a legacy database and a few hundred custom hooks rather than a clean demo install.
I have followed the WordPress Core AI evolution since the first building blocks landed, and the tracing work (LangSmith, Langfuse) matters more than it sounds. Without a trace you are guessing about why a post-management ability quietly did nothing.
If keeping up with the WordPress 7.0 AI features is eating your dev hours, I have been doing this since the 4.x days and I am happy to take it off your plate.
Where the 7.0 AI work stands
- Beta 1 lands February 19th, four weeks out.
- New filters will let admins limit AI access by context and user role.
- The PHP AI Client will most likely follow the
Requestslibrary pattern into core. - WP Bench is turning into the default way to score model performance on WordPress tasks.
The Make/Core 7.0 roadmap is where the official updates land, and the team’s AI vision posts fill in the reasoning behind them.
