BetterShield
Appearance
MCPAI assistantsAgent activity

What can an AI assistant read about your WordPress security?

· · 8 min read

An AI assistant that you connect to BetterShield with read access can see your site’s security information, such as the score, the findings, the available fixes and recovery readiness, and it cannot change anything. Reading is a single switch, Let AI assistants see security info, and it is off until you turn it on.

This post is about the reading side only. If you want to connect an assistant first, WordPress MCP: connecting an AI assistant to your site safely walks through it, and the change side, with approvals and undo, is in WordPress MCP write access: approvals, limits and undo. Here we cover what a read returns, what it never returns, and how to see for yourself what your assistant is allowed to look at.

Quick summary

  • Reading is one switch, Let AI assistants see security info, under BetterShield › Agents › Permissions. It is off by default.
  • With it on, BetterShield offers twenty-eight bounded, read-only abilities. None of them changes the site.
  • One of them, get-security-overview, returns the score, its history, the findings with available fixes and recovery readiness in a single read.
  • Credentials, connector keys, file contents, named alert recipients and usernames in lockout data are never returned.
  • Reading is separate from changing. Turning reading off also turns change access off.

Why it helps to let an assistant read

A security screen is full of facts, and most of them need a little translation. An assistant that can read the same facts can answer a plain question such as “what should I fix first?” without you opening five tabs. Because the answer comes from the site itself, it describes your site, not a typical one.

Reading is also the low-risk half of the arrangement. An assistant that can only read cannot alter a setting, close a door you meant to leave open or lock anyone out. That is why we suggest starting here and staying here until you have a reason to go further. The principle of least privilege is the idea behind it: give a tool the smallest set of powers that gets the job done.

What a read returns

When reading is on, BetterShield registers its readable abilities with WordPress. In the plugin’s own words, they can report scores, findings, protections, selected settings, activity, incidents, file-check status, versions, connection state and related security information. In plain terms, an assistant can answer questions in these areas:

AreaWhat an assistant can learn
Score and findingsThe audit score, what is open, and the fix options a finding has
ProtectionsWhich protections are on, and selected settings
Activity and incidentsWhat happened on the site, and the incidents recorded
FilesThe status of the file check
UpkeepVersions, and the state of the assistant connection itself

The list in your own install also depends on two things the plugin tells you about: the permissions you have turned on, and the tools you have chosen for each app. Limit what each AI assistant can do covers the second. An app limited to a few tools sees only those.

The security overview

The release that added get-security-overview describes it as a tool that reads the score, history, findings with fixes, and recovery readiness. It is the best first question for an assistant to ask, because one read answers most of what a site owner wants to know: how healthy the site is, whether that is getting better, what is open, and whether you could get back in if a sign-in change went wrong.

Recovery readiness is the part people forget. BetterShield checks that you have a way back in before it asks you to tighten sign-in, and an assistant that can see that state can tell you when it is the weak spot. A second read lists the fix options on a finding, so an assistant can look at what is possible before anyone proposes a fix. Vulnerability data appears only once you have turned on vulnerability checks.

What a read never returns

This is the half that matters most for trust. Sensitive values stay out of every read: credentials, connector keys, file contents, named alert recipients and usernames in lockout data. An assistant can learn that a file changed. It cannot read what is in the file.

Some actions are also never possible through any connection, whatever you switch on. Creating accounts or credentials, changing recovery options, two-factor or alert settings, lifting lockouts and deleting log rows are all out of reach. Reading adds none of them.

Last, BetterShield never ships, asks for or stores an AI key for this. The assistant is yours: you add BetterShield to Claude, ChatGPT or an editor, and that connection is what reads.

Set it up

  1. Open BetterShield › Agents › Permissions and turn on Let AI assistants see security info. It is off by default.
  2. Or open BetterShield › Agents › Connect and turn on Allow AI assistant connections, which also turns reading on.
  3. Under Add BetterShield to your assistant, pick your assistant. For Claude or ChatGPT, press Copy the Address, add it to the assistant and sign in through your browser. For an editor, press Set Up an Editor and Create Key.
  4. Leave Allow this key to make fixes unticked, so the key can only look.
  5. Press Test Connection under Connection diagnostics. It makes one real call and shows how far it got.
  6. Ask your assistant a plain question, such as “what is my security score and what should I fix first?”

Turning reading on does not connect anything by itself. A connection still needs its own key or browser approval, and that is the second lock.

Try these questions

Read-only access suits questions, and these are good first ones. None is a product claim, only a prompt you can adapt:

  • “What is my score now, and how has it changed?”
  • “Which open findings have a fix I can undo?”
  • “Is my recovery link ready, in case I change sign-in settings?”
  • “What happened on the site this week?”
  • “Which plugins on this site offer abilities to assistants?”

If an answer sounds wrong, open the matching screen and compare. The assistant reads what BetterShield shows. It does not know more than the screen does.

What it can break, and how to undo

Read abilities do not change the site, so there is nothing to roll back. What reading does is let an authorized connection see your security posture, which is why we suggest connecting only apps you use.

To take it back, turn Let AI assistants see security info off. That removes BetterShield’s abilities and also turns agent change access off. To end one connection, press Disconnect on its row under Connected apps. Disconnect All ends the key and every signed-in app.

Check it is working

Three places confirm the state:

  • The AI assistant access strip at the top of the Agents screens shows whether reading is on.
  • The registered abilities list, on the Surface tab under BetterShield › Agents, shows the abilities on your site. Turning reading off removes BetterShield’s from the list. The Abilities API handbook explains what an ability is, and WordPress Abilities API: see what AI agents can reach and do shows how to read the list.
  • Activity › Agent activity records every call, including refused ones, with the outcome and a digest of the input, never the input itself. Record what agents do on this site is on by default.

A good test is one read your assistant is allowed to make and one it is not. Both should show up, with different outcomes.

On a multisite network, access is set per site, and a super administrator can read the limited network ability where available.

We recommend starting with reading only: one dedicated connection per assistant, Allow this key to make fixes unticked, and activity recording left on. Review the registered abilities list after you install or update a plugin that offers abilities. If you later want an assistant to act, read Let an AI assistant apply approved changes first, and keep the first change small.

Frequently asked questions

Can a read-only AI assistant change my WordPress site?

No. With only reading on, an assistant can look and change nothing. Changing is a separate switch, Let AI assistants fix things, and it is off by default too.

Does the assistant see my passwords or keys?

No. Credentials, connector keys, file contents, named alert recipients and usernames in lockout data are not returned. BetterShield also never stores an AI key.

How many things can an assistant read?

BetterShield offers twenty-eight bounded, read-only abilities. Which ones a given app can use depends on the permissions you have on and the tools you chose for that app.

What is the security overview?

It is one read that returns the score, its history, the findings with available fixes and recovery readiness, so an assistant can summarize the whole picture at once.

How do I see what the assistant looked at?

Open Activity › Agent activity. Each row shows the ability, the outcome and a digest of the input, and refused calls are included.

Conclusion

Letting an assistant read is the gentle way to start: it gets answers from your own site, and it can change nothing. Turn on reading, connect one app with a key that can only look, run Test Connection, and check Agent activity to see what it asked for. Add change access later, only if you have a reason.

BetterShield is free on WordPress.org. The doc for this feature is Let an AI assistant read security information, and Review abilities registered on this site explains the list.

Close the open doors today

Install the free plugin. The first audit runs when you activate it, and all but one of the fixes wait for your choice.

Requires WordPress 6.7 or newer and PHP 8.0 or newer.

Get BetterShield