Changelog do AI Engine: ChatGPT Chatbot, GPT Content Generator, Custom Playground & Features (Pro)

Correção

• Security: Uploads are now validated against their actual content and not only their filename, so HTML disguised as audio or another media type is refused in chat uploads, form uploads, the Public API and generated files (reported by peppa via Patchstack).
• Documents sent to a chatbot on Azure OpenAI now upload to Azure itself, and PDFs work there.
• A full API address pasted as the Azure endpoint is now trimmed to the resource URL.
• Claude can now connect as a custom connector on hosts that answer /.well-known/ themselves, such as SiteGround.

Correção

• Chatbots with image generation work again when a reasoning level is set.
• Knowledge sync marks a post as failed when the model returns no vector, instead of crashing and leaving half-written rows.
• Gemini 3.1 Flash Image is billed at its real price instead of half.

Correção

• Security: Image Studio now encodes the media ID in its requests, as extra hardening.
• Trashing a post now deletes its vectors through each environment's own vector DB.
• On mobile, the Glass and Timeless chatbots keep the input above the keyboard in long conversations.
• A page that places the site-wide chatbot as a popup no longer shows it twice.

Correção

• Security: Shop Managers can no longer edit administrator accounts through MCP, and the plugin and theme file tools now follow DISALLOW_FILE_EDIT.
• Security: Forms and the Public API can no longer read protected uploads, the Google voice chatbot no longer exposes its API key, voice chats cannot replace someone else's conversation, and only admins can read the logs.
• Security: Guests can no longer reset their usage limits with a fake IP header, while proxies and Cloudflare are still recognized.
• Security: Chatbot replies now drop embedded frames and forms, and guest uploads are capped at 20 per hour.
• Chatbots keep web search, MCP servers and their reasoning settings after calling a function, which also stops Claude's "Invalid signature in thinking block" error.
• Chatbots keep their Knowledge context after calling a function on every provider, including Google, OpenRouter and other compatible providers.
• Syncing a large OpenAI Vector Store no longer times out, each step stops after about 25 seconds and the next one continues from there.
• The chatbot preview in the Chatbots tab shows the conversations held before 3.8.7 again instead of starting empty.
• Content Studio no longer shows each section heading twice when the model repeats it.
• MCP tools now accept text request ids as the JSON-RPC spec allows.
• Re-setting a featured image through MCP no longer fails, terms can be removed, and featured image errors say why.
• Admin scripts refresh when only the UI library changes instead of staying cached.

Correção

• Security: Voice chatbots now only run functions attached to their own chatbot.
• Security: Editors on MCP can no longer publish other content types or file references through a decoy fields object, and Public API tools and MCP servers stay restricted to administrators.
• Security: Contributors can no longer take over a live chatbot through WordPress' embed preview, and outside post content a shortcode can no longer replace a live chatbot's prompt, model or API key.
• Security: The Gemini file upload now verifies TLS.
• Security: The MCP sign-in page can no longer be framed by another site, shows where the app sends you, and warns when an app calls itself Claude or ChatGPT without being it.
• Insights logs work again on sites whose logs table predates 2.5.6, since tables now get their missing columns on update.
• Chatbots and forms no longer stay blank when one of their settings contains entity text like ".
• Posts containing an escaped "<" (like "<48 dB") no longer lose everything after it in Knowledge and Content-Aware.
• AI Forms in chatbot mode now use the chatbot's own Max Tokens, model and temperature instead of hidden form values that cut replies at 4096 tokens.
• Messages, uploads, stopping a reply and loading a conversation now follow the right chatbot's discussion when a chat id belongs to another chatbot.
• Claude no longer rejects the answer after a function call when Knowledge is on (Invalid signature in thinking block).
• When a thinking model spends its whole Max Tokens reasoning, the reply now says so instead of showing the raw reasoning as the answer.
• Knowledge no longer retries rows of deleted environments, and large OpenAI Vector Stores sync page by page instead of timing out.
• MCP connections no longer drop when a client revokes the token it just replaced after a refresh, and sign-in responses say which site answered.
• Links in the GDPR consent text are now clickable instead of shown as raw text.
• The site-wide popup chatbot no longer leaves a screen of blank space at the bottom of the page on themes that give the body a height.
• The user's IP is shown again under their name in Discussions and Insights for logged-in users, and in the discussion details.
• Chatbot names with & no longer show &.
• Admin screens no longer load blank, and the header now shows the Free and Pro tiles.
• Links on other plugins' screens are no longer tagged as AI Engine traffic.
• OVHcloud and Sakana AI now follow one rule: hidden until switched on with MWAI_OVH or MWAI_SAKANA, and always shown where an environment of that type exists.

Correção

• Security: Editors can no longer change the title or excerpt of posts they are not allowed to edit (coupons, menus, templates) through the AI helpers, and the other helpers taking a post ID now check permissions too (reported by Anatoly Gorbounov, x1will, via WPScan).
• Hotfix: Chatbot replies no longer crash on servers with an older cURL, such as Xserver, because of a constant added in 3.8.5.
• Messages like "0" or an image alone no longer fail the reply when Knowledge is on.
• Chatbot saves silently refused by the database now show an error.
• A Max Height set without a unit is now treated as pixels.
• AI Forms set to a specific environment with the Default model now work instead of failing with "Model not found".
• Settings changed while a chatbot reply is running are no longer undone when that reply saves its usage stats, which could revert a refreshed model list.

Correção

• Security: MCP no longer returns API keys or tokens when reading WordPress options.
• Security: Pro MCP tools now refuse accounts that cannot manage the site instead of running as the first user, and they pause when the license lapses.
• Security: Uploaded files can no longer carry metadata AI Engine uses internally, and the expired files cleanup only deletes the expired file itself.
• ChatGPT connections no longer fail with a missing resultType error when MCP requests use the 2026-07-28 format.
• ChatGPT and Claude can now sign in to MCP on servers that block paths containing /.well-known/ under /wp-json/, like many nginx setups.
• Connecting ChatGPT or Claude no longer fails with an expired code right after approval on sites with a persistent object cache.
• Claude Sonnet 5.5 cache reads now cost $0.10, as Anthropic charges.
• AI Engine no longer loads the block editor scripts on Classic Editor screens, which broke plugins like HTML Editor Syntax Highlighter.

Correção

• Security: A chatbot shortcode in a post can now only change the prompt, model or API key when its author can edit others' posts.
• Security: The search endpoint used by the Search settings screen is now limited to administrators.
• Qdrant and Chroma server URLs pasted from the browser, like :6333/dashboard#, now work instead of sending every request to the dashboard page.
• Streamed replies are no longer cut after 5 minutes while still arriving, only when the connection goes silent.
• The JSON helpers used by Meow Apps Health and add-ons now work on Claude and Gemini sites with default settings.
• Uploading to an OpenAI Vector Store now says exactly what is wrong with the environment, such as a missing store ID, no OpenAI environment or no API key.
• Pro sites whose server cannot reach the update server no longer hit a fatal error on PHP 8 when WordPress checks for updates.

Correção

• The Editor Assistant no longer fails with Feedback session expired after applying edits, saved queries no longer store API keys, and it now sees paragraphs edited after the page loaded.
• Claude, Gemini and other non-OpenAI chatbots on an OpenAI Vector Store now receive the matching knowledge instead of an empty context.
• Non-streamed errors from OpenRouter, Mistral, OVH and custom providers are shown as a message instead of raw JSON.
• Gemini replies cut by Max Tokens end with the notice instead of an empty-reply error, and Gemini errors wrapped in a list are readable.
• Picking an environment without a model now asks for one instead of sending another provider's default model and failing with a 404.
• An AI Form repeated on a page answers in its own copy, and a streamed form error no longer floods the PHP log.
• GPT-6 Astra can generate images, its model entry was missing the image generation tool.
• Uploads that fail now say why, with the size limit when the file is too large, instead of an unknown error or a wrong invalid file type message.
• An empty completion request is refused before reaching the provider, and Suggest Synonyms without a selection returns a clear error instead of a fatal.
• MCP tool failures, like a missing post, reach the AI as a readable tool error instead of a protocol error.
• ChatGPT MCP connections no longer drop when several of its sessions share one connection, and revoking an app disconnects all of its sessions at once.
• When AI Engine manages the WordPress Connectors, an empty key no longer wipes a saved one, and keys already saved in WordPress are reused.
• Syncing a single post that gets skipped (ignored, language filter, or filtered out) now says why instead of doing nothing.
• Copying a masked MCP command by hand copies the real token, and Knowledge asks to pick an environment instead of showing the first-run screen.
• On touch screens the Stop button works with a single tap, while a quick double tap on Send still cannot cancel the message.
• A failed license check shows the real reason straight away, and a successful retry no longer keeps saying the Pro version is disabled.
• Two PHP warnings on every logs request when Insights is off.

Correção

• A site that cannot reach the license server no longer hangs the admin for minutes: connection diagnostics run once instead of once per URL, and the license check times out after 20 seconds instead of 45.
• MCP connections no longer break when a client, such as ChatGPT, reuses a refresh token that was already replaced.
• Chatbots with streaming disabled no longer slide the window up when a reply starts, which hid the Glass header and could make the page jump.
• Add-ons now keep their own visitor params as sent; only denied keys are matched by their canonical name and only messages is folded into one key.

Correção

• Security: visitors could override a chatbot's or form's model, instructions or environment with renamed parameters. Fixed, and the Advisor widget now escapes its output.
• Embeddings were skipped on pages made with a page builder (Breakdance, Oxygen, Bricks...).
• Models of a Custom (OpenAI-compatible) environment could not receive images.
• The discussions list with several lists on one page, empty discussions, and its menu while scrolling.
• The admin crashed when a chatbot's ID was changed, and a crash in one settings tab no longer locks the whole screen.
• Stopped or edited Workspace replies no longer come back duplicated.
• A streamed reply was rebuilt when it finished, reloading its images.
• Dev Tools no longer polls the server every second while open.

Correção

• Daily, weekly and monthly limits in Absolute mode no longer trip early.
• Image, media and post creation routes now check permissions before acting.
• Listing chatbots no longer fatals on the free plugin when a chatbot uses functions.

Correção

• Workspace module no longer switches itself off on sites without a Pro licence.
• OpenAI Vector Store no longer retrieves knowledge twice on the default environment, now passes Max Results to file_search, and repairs old vectors tables missing columns.
• Uploaded files stay in context on later turns with Claude and other stateless providers.

Correção

• Pairing failures that consumed the token now ask the user to generate a new code.
• Connection test no longer always reports success.
• Gemini no longer offers Live-only models as chatbots.
• Function calling on Google's Standard API.
• AI Copilot no longer stretches paragraphs to the full editor width.

Correção

• Temperature is no longer sent to models that reject it.

Correção

• Anthropic now caps max_tokens to the model's limit instead of returning a 400 error.
• Provider errors with a flat body (such as Mistral) and any 4xx/5xx status now show the actual error message.

Correção

• Sending a message after a client-side JS function call no longer fails with an invalid_request_error.
• MCP OAuth login redirects are no longer cached by page caches and served to the wrong visitor.
• Plugin file tools no longer report a refusal as a missing directory.

Correção

• Chatbots using the default environment and model are no longer excluded from the function calling list.
• Vision now detects image types from the file contents instead of the extension, so images served through a CDN work.
• Magic Wand can read the selected text again in the iframed block editor of WordPress 7.1, re-enabling Suggest Synonyms.

Correção

• Editor Assistant endpoints accepted a nonce any visitor could generate, allowing anyone to run AI queries on the site's provider account (reported by Abdullah Kareem via WPScan).
• A crafted URL could escape the uploads folder and send any server-readable file to the AI provider (reported by Jashid Sany via WPScan).
• Editor Assistant now works for admin-equivalent accounts whose role isn't literally administrator.
• Embeddings sharing the same reference no longer stay stuck, and a single unprocessable vector no longer blocks the maintenance queue.
• A temperature of 0 is no longer treated as unset and stripped from the request.
• Image-only replies from GPT-5 models no longer show "medium" as the answer text.

Correção

• The mwai_mcp_mutate hook now fires on deletions, so cache purges hooked to it run.

Correção

• Qdrant collections were always created with 1536 dimensions regardless of the embedding model, and default dimensions are now read from the right place.
• The MCP access level no longer restricts OAuth connections, and the setting is no longer hidden behind the bearer token.
• MCP connections no longer drop at random when the client refreshes its access token.

Correção

• Workspace mobile app connection on servers that don't expose the Authorization header, with clearer pairing errors.
• A Context Max Length of zero no longer empties posts, which blocked embeddings sync and blanked chatbot context.

Correção

• Public chat and form endpoints no longer trust client-supplied models, environments, instructions or messages, and internal errors are hidden from visitors; the stateful-response restore is now owner-scoped, so another user's chat ID can no longer resume their conversation.
• Content-aware chatbots no longer inject a post the visitor cannot read, so a client-supplied context ID can no longer leak draft or private content.
• Chatbot error and notice messages are sanitized before being rendered as HTML, and configuration errors are shown to editors only instead of being printed publicly.
• Function calling is more reliable: Claude no longer re-calls tools with the previous message's parameters, multi-round calls on stateless providers keep their feedback blocks instead of clearing them each round, and the Gemini 2.x models work again.
• Google model names and capability detection, so image, thinking and web-search models are recognised and preview variants are no longer ambiguous.
• OVH replies arriving empty, caused by an oversized max_tokens and a hijacked catalog fetch.
• Embeddings re-sync destroying a vector when the new content comes back empty, and guest file listing crashing with a 500.
• The usage-limit lock never firing on streaming chatbots, and the Discussions table spinning forever when opened in a background tab.

Correção

• Fixed block editor paragraph spacing broken by the AI Copilot handler wrapper.
• Fixed the Anthropic engine dropping conversation history on function-call follow-up requests.
• Fixed duplicate embedding vectors for the same reference and environment, which left embeddings stuck at pending after a sync.

Correção

• MCP user tools not enforcing proper capabilities, allowing Multisite site administrators to edit users beyond their reach. Hotfix. (Thanks to Charles Vosburgh for reporting.)
• MCP read tools corrupting block-JSON Unicode escapes by double-unslashing database content.

Correção

• Gemini image models not being selectable when their ID lacks a -preview suffix.
• MCP media upload tool crashing on download errors and rejecting extensionless image URLs.

Correção

• MCP post meta tools double-serializing array values, which stored nested arrays as strings.
• MCP wp_create_post and wp_update_post tools stripping shortcodes, email HTML, and inline CSS for admin-authorized callers.
• Chatbot theme styles loading outside AI Engine's admin screens, which caused Gutenberg editor iframe warnings.
• simpleJsonQuery failing on Anthropic and Google environments by no longer forcing an OpenAI fallback model, with a clearer error message.
• Discussions and Queries timestamps now display in the site's configured timezone instead of the browser's.
• MCP Functions settings crashing when tools were registered due to the list being returned as an object instead of an array.
• OpenAI Vector Store Pull All failing with a limit error by paginating the file list.

Correção

• Code blocks now display correctly in the Timeless and Foundation chatbot themes (was using proportional font instead of monospace).
• Sonnet 5 breaking with "each thinking block must contain thinking" when streaming with function calling.
• MacOS chatbot container showing empty notches in the corners due to double-rounded borders.

Correção

• wp_create_post MCP tool no longer corrupts shortcode and HTML content via the markdown parser.
• MCP log entries mislabeled as OAuth; admin and bearer calls now display correctly.
• Streaming parallel tool calls leaving a null in the tool_calls array, which broke function calling on strict providers.
• MCP wp_get_posts, wp_get_media, and wp_get_comments filters by author, user_id, and email were accepted but silently ignored.
• wp_update_option now decodes JSON-string array values into real arrays before saving, preventing corruption of array options like sticky_posts.
• Chatbot events panel is now scrollable and error messages are dismissable.

Correção

• Add-ons were reported as disabled whenever options were fetched outside the admin context.
• Max Tokens entered with a thousands separator (e.g. the Danish "64.000") was read as 64 and capped responses.
• Content Aware placeholders now resolve automatically in chatbot instructions without requiring a toggle.
• Setup Assistant no longer shows "Connect an AI Provider" as incomplete when only the Fast model is deprecated.
• Realtime instructions now resolve placeholders like {FIRST_NAME} into their values.
• MCP tools generated from WordPress REST endpoints could break Google Gemini due to enum on non-string types.
• Changing the type of an AI environment now drops fields from the previous type (endpoint, region, deployments), preventing engines from calling the wrong provider's endpoint.
• The "Default AI Environment not configured" warning no longer appears when a Chroma embeddings environment generates its own vectors.
• The Usage widget on the Dashboard no longer overflows its panel when numbers are wide.
• Anthropic's streaming ping keepalive events are now ignored instead of being logged as an unknown stream data type.

Correção

• Hold to Talk button on mobile Realtime chatbot no longer triggers the context menu (copy, select) when held.
• Sending is now blocked while a file is still uploading, with a spinner on the send button and a progress bar on the file preview until upload completes.
• Eliminated the gap between the chatbot and the keyboard on iOS by locking body scroll while the mobile chatbot is open.
• Anthropic safety refusals now show a readable message instead of an empty reply.
• OpenClaw link in the readme now points to our own article.
• PHP 8.2+ deprecation warning caused by an undeclared $extraData property on Meow_MWAI_Reply.
• wp_alter_post regex mode now handles patterns containing "/" (e.g. Gutenberg block markers) and returns the actual PCRE error instead of a generic one.

Correção

• Filenames with underscores (e.g. Some_File_2026.pdf) were getting partially italicized by the chatbot's markdown renderer.
• ReferenceError (data is not defined) in the chatbot response handler that masked the real error message when a host returned a non-JSON body.
• The Push All button showed "(null)" when a post type had zero matching posts.
• The wp_update_option MCP tool schema no longer logs a "problematic union type" warning on every list_tools call.
• The Meow Apps menu icon now shows correctly in both the expanded and collapsed sidebar states.
• Hotfix for the MCP endpoint URL being hidden until a bearer token was generated, even though OAuth doesn't need one. It now appears as soon as the MCP Server is enabled.