July 14, 2026 — WordPress AI 1.1.0 shipped June 30 with two experiments and a quieter theme running underneath them: after 1.0.0 gave publishers the audit trail to see what their AI was doing, 1.1.0 starts governing it at the edges. Key Encryption moves connector credentials behind encryption at rest, Type-ahead Text brings ghost-text drafting into the block editor, and a set of readiness gates and moderation opt-outs cut the wasteful requests that inflate provider bills. For newsrooms running AI in production, this is the release where the platform stops asking “can we?” and starts asking “should this request run at all?”
Key Themes
- Key Encryption stores AI Connector API keys encrypted in the database and decrypts them only when needed, closing a credential-at-rest gap that security reviews routinely flag.
- Type-ahead Text adds inline ghost-text suggestions at the end of paragraph blocks — the editor-facing feature publishers have compared to Google Docs and Notion — and it is Block Editor-only.
- Content-readiness gates now disable Resizing, Summarization, and Classification until a locale-aware word or character count is met, so the tools stop firing on thin content.
- Comment moderation gains a switch to stop analyzing guest comments and skips anything already marked spam or trash, turning moderation spend into an editorial decision.
- A new
core/read-settingsability and thewpai_has_image_generation_supportfilter continue the Abilities API’s expansion as it marches toward core.
Jump to: 🔐 Key Encryption · ✏️ Type-Ahead Text · 🚦 Request Discipline · 🛠️ Abilities API · 🔮 What’s Next · 💡 Takeaways
🔐 Key Encryption
The first of two experiments in 1.1.0, Key Encryption encrypts AI Connector API keys before they are written to the database and automatically decrypts them at the moment of use. If the experiment is disabled, the functionality is turned off, or the plugin is deactivated, the original keys are restored — so the encryption layer is reversible rather than a one-way migration that could strand a site’s credentials.
This is unglamorous plumbing, and that is exactly why it matters. Provider API keys have historically lived as plaintext in wp_options, readable by anyone with database access, a stray backup, or a debug dump.
Why this matters: For publishers, an OpenAI or Anthropic key is not a config value — it is a spend credential and a data-access grant rolled into one. A leaked key means someone else’s inference on your budget, and potentially your content flowing through their prompts. Security reviews at enterprise media companies flag plaintext secrets at rest as a standard finding, and “the CMS stores provider keys encrypted” is the kind of sentence that clears a vendor assessment instead of stalling it. For WordPress VIP clients and other managed customers whose procurement teams run formal security questionnaires, Key Encryption turns a recurring objection into a checkbox. It pairs naturally with the connector approvals that landed in 1.0.0 — approvals decide which plugins may use a credential; encryption protects the credential itself.
✏️ Type-Ahead Text
The second experiment is the one editors will actually see. Type-ahead Text surfaces inline ghost-text suggestions at the end of paragraph blocks as a writer works, with manual triggering inside supported blocks, context-aware continuations, and keyboard controls for accepting or dismissing a suggestion. Developers can override the provider and model through developer settings, and — critically — the feature integrates with the site’s Guidelines (Knowledge), so completions can be steered by house style rather than a generic model default. Documentation now states plainly that this is Block Editor-only.
Why this matters: Ghost text is the most visible AI surface a newsroom can hand a reporter, and it is also the one that most directly touches editorial accountability. Because Type-ahead can draw on site Guidelines, a publisher can constrain suggestions toward its own voice and standards instead of accepting whatever the model volunteers — a real governance lever, not a cosmetic one. But the same feature writes machine-generated phrasing directly into the copy, which raises the question every editor-in-chief will ask: how do we know which sentences a human wrote? The completion is invisible once accepted. Publishers piloting Type-ahead should decide up front how AI-assisted drafting is disclosed internally, and watch the C2PA text-provenance work on the roadmap — the two features are on a collision course, and provenance is the answer to the trust question ghost text creates.
🚦 Request Discipline
Underneath the experiments, 1.1.0 spends real effort on not calling the AI. Features that require post content — Content Resizing, Content Summarization, and Content Classification — are now disabled until a minimum threshold is met, measured with locale-aware word or character counting so the gate works for languages where words are the wrong unit. On the moderation side, a new setting controls whether guest comments receive automatic analysis at all (it defaults to enabled), and the moderator now skips comments already marked spam or trash, so the site stops paying to analyze content it has already rejected. The AI Request Logs window is also fixed at a consistent 30 days.
Why this matters: Every one of these changes is a line item on a provider invoice. In 1.0.0, request logging gave editorial operations teams visibility into which providers were processing content and how often — the answer to “why did the OpenAI bill triple?” In 1.1.0, WordPress starts acting on that visibility: no summaries of single paragraphs, no classification runs on empty drafts, no toxicity scoring on comments already headed for the spam folder. The locale-aware counting is the detail multilingual publishers should notice — a threshold measured in words breaks for Japanese, Chinese, or Korean, and this release counts characters where that is the right measure. Together these are the guardrails that keep AI spend proportional to editorial value, which is the difference between a pilot that survives budget review and one that gets cut.
🛠️ Abilities API and Developer Surface
The Abilities API keeps expanding. 1.1.0 adds core/read-settings, a read-only ability that exposes eligible site settings through the Abilities API and requires appropriate permissions to invoke — permissioned, least-privilege reads rather than blanket option access. A new wpai_has_image_generation_support filter lets third-party plugins declare image-generation capability. The Abilities Explorer itself gets more usable: cleaner rendering of non-ASCII characters, better handling of schema defaults in ability inputs, and support for scalar input-schema validation.
Why this matters: Platform teams building house AI tooling on top of WordPress are the audience here, and the direction is toward the same governance posture the user-facing features show — reads are scoped and permissioned, not open. This is the Abilities API’s steady march into core covered in this month’s developer roundup, now with a concrete read-settings primitive that internal tools can build against.
🔮 What’s Next in 1.2.0
The 1.2.0 milestone lines up real-time collaboration for Refine-Notes, a Suggest Replies experiment, a Content Translations experiment, Gutenberg Media Editor integration with crop suggestions, promoted AI Provider Connectors, and — the item publishers have been waiting on — C2PA content provenance for both text and images.
Why this matters: Worth a note on the arc. The 1.0.0 recap flagged C2PA and an AI Playground as headed for 1.1.0; neither landed here — C2PA has slipped to 1.2.0 and the Playground is absent, while Type-ahead arrived as predicted. That is a normal roadmap adjusting to reality, but publishers building an AI-disclosure posture around platform-native provenance should plan for the later date, not the earlier promise. Content Translations is the sleeper item for international media groups, and Suggest Replies extends AI from moderating comments to drafting responses in them — a bigger editorial-accountability question than moderation ever was.
🔧 Also in 1.1.0
Editorial Notes now prompt to save after a run, closing a lost-work gap. Loading indicators are scoped to the active block instead of the whole editor, and Editorial Updates appear only when they are linked to current blocks. Title Generation gets consistent button text and loading states, and Content Classification term suggestions now persist reliably rather than flickering away. Experiment modal warnings moved to the standard Notice component for a more native feel, and documentation was clarified to state that AI features are Block Editor-only. The release spans 111 commits across 213 files with 30-plus contributors between 1.0.0 and 1.1.0.
💡 Takeaways
- Audit your credential storage and enable Key Encryption as soon as it reaches your environment — plaintext provider keys are a standard security-review finding, and this is the fix that clears the questionnaire.
- Pilot Type-ahead Text in a controlled newsroom setting with your site Guidelines populated, and decide how AI-assisted drafting is disclosed internally before reporters rely on it.
- Define content-readiness thresholds deliberately — the locale-aware gates only help if the minimums match how your desk actually writes, and multilingual teams should confirm character counting behaves for CJK content.
- Evaluate the guest-comment moderation switch against your spend — turning off analysis for anonymous or already-flagged comments is now a lever you control, and it maps directly to the request logs from 1.0.0.
- Watch the C2PA slip to 1.2.0 — provenance is the trust infrastructure that answers the questions Type-ahead raises, and it is now a quarter later than the earlier roadmap implied.
The full release post is available on Make WordPress AI.
Update (July 23, 2026): AI 1.2.0 shipped Suggest Replies as predicted, but C2PA provenance slipped again — from 1.2.0 to the 1.3.0 milestone. See the WordPress AI 1.2.0 recap.