Lock the Doors: AI Security Basics for a Small Office · Lesson 4 of 7

Connectors, agents and the accounts behind them

20 minutes. Read-only first. Approvals for anything that sends, pays or deletes. Two-step sign-in on every account.

---

Lesson 3 was about what an AI app can reach on a computer. This lesson is about what it can reach in your accounts: your email, your files, your accounting, your customer list.

Two words first.

  • A connector links an AI assistant to another service, like Gmail, Outlook, Google Drive, QuickBooks or HubSpot, so it can look things up there, and sometimes make changes.
  • An agent is an AI tool that can carry out several steps on its own: read your inbox, draft replies, update a record, file a document.

Both work the same way underneath. The AI uses an account's sign-in. It can see and do whatever that account can see and do. So the safety of a connector comes down to three things: what it's allowed to do, whether a person approves the risky parts, and how well the account behind it is protected.

Habit 1: Read-only first

Every new connector or agent starts by reading only, for at least a week. Watch what it does. Learn where it's reliable. Then allow it to draft things, one tool at a time.

In Claude, for example, each connector has its own page under Customize, then Connectors. Under Tool permissions you can set each group of tools to Always allow, Needs approval, or Blocked. Set anything that writes, sends or deletes to Needs approval or Blocked. ChatGPT and Microsoft 365 Copilot have similar controls; the course Connect Your AI to Your Business Tools walks through all three in its Lessons 2 and 6.

Habit 2: A person approves anything that sends, pays or deletes

Anything that leaves the building or can't be undone needs a person to say yes, every time:

  • Sending an email, message or invite
  • Paying, refunding, invoicing or changing a price
  • Deleting records or files
  • Changing who can see a file

When an approval prompt shows up, read it. Check who it goes to, what it says or changes, and how much if money is involved. If you didn't expect the prompt, say no. An unexpected request to send or share is the classic sign that the AI read hidden instructions in an email or web page and is trying to follow them.

And one more for agents: if an agent can act on the web or in your accounts, it should keep a record you can read and it can't erase. A forensics firm reported in early October 2026 that AI agents it studied used throwaway accounts and, in some cases, erased records. Whatever the reason, the lesson for an office is simple: no record, no permission to act.

Habit 3: Know the off switch, at both ends

Practice turning a connection off before you need to.

From the AI side:

  • Claude: in a conversation, select + in the message box, then Connectors, and switch it off for that chat. To sign Claude out of the service completely, go to Customize, then Connectors, select it, and choose Disconnect.
  • ChatGPT and Copilot: look in the app's settings for connected apps or sources. The names change often; the Connect Your AI course has the current paths.

From the account side (the surest off switch, because it cuts the connection from the service's end):

  • Google account: go to `myaccount.google.com/linkedapps`, pick the category (for example, apps with Access to your Google Account), select the app, choose See details, then Remove access.
  • Microsoft 365 work account: in the My Apps portal, hover over an app, select the ... menu, then Manage your application. The top section shows what you personally allowed, with a Revoke Permissions option. The bottom section shows what your administrator allowed for the whole company; only the admin can change those.
  • Business apps (QuickBooks, HubSpot, your store and so on): look in the app's own settings for "Connected apps," "Integrations" or "Security."

For the whole company (admin or IT partner):

  • Google Workspace: in the Admin console, Security, then Access and data control, then API controls, then Manage App Access. It lists the outside apps that have reached your company's Google data and how many people use each. Each one can be set to Trusted, Limited, Specific Google data, or Blocked.
  • Microsoft 365: in the Microsoft Entra admin center, Entra ID, then Enterprise apps, then All applications. Select an app, then Permissions, to see what it was granted for the whole company and for individual people.

You don't have to do the company-wide check yourself. Ask your IT partner to send you that list. It doubles as your first sweep for unofficial AI tools in Lesson 5.

Habit 4: Two-step sign-in on every account an AI tool touches

Multi-factor authentication (MFA), also called two-step sign-in or 2-Step Verification, means a password alone isn't enough. You also need a second proof, usually a code or a tap in an app on your phone. If someone steals the password, they still can't get in.

It matters double for AI. An account that a connector uses is an account with your AI's reach attached to it. Turn on MFA for:

  • The AI tool's own account (your Claude, ChatGPT or Copilot login).
  • Every account it connects to: email, files, accounting, CRM, payments.
  • Every admin account, first and above all.

Where to turn it on:

  • Microsoft 365, for the whole company: Microsoft offers free security defaults, which require everyone to register for MFA, require admins to use it, and block older sign-in methods that skip it. An admin turns it on in the Microsoft Entra admin center under Entra ID, then Overview, then Properties, then Manage security defaults. Microsoft notes that companies with more advanced licenses often use Conditional Access policies instead. If you have an IT partner, ask which one you're on. Either is fine; "neither" is not.
  • A personal or single-user Google account: at `myaccount.google.com`, select Security & sign-in, then Turn on 2-Step Verification. For a Google Workspace work account, Google says these steps may not work and to ask your administrator, because the company controls it.
  • Business apps: look in each app's security or profile settings for "two-step," "2FA" or "multi-factor."
Do this now: one row per connection

Open your Front-Door List and go to Part 2 (AI tools). For each AI tool:

  • List what it's connected to (email, files, accounting, and so on).
  • Write its mode: Read-only, Read and draft, or Can act. Anything at "Can act" without approvals gets changed today.
  • Mark MFA on for the AI tool's account and for each account it connects to. Turn it on where it's missing.
  • Find one connection nobody uses and turn it off, from the account side. That's your off-switch practice.

Check it worked

  • Every AI tool on your list shows what it's connected to and its mode.
  • Nothing sends, pays or deletes without a person approving it.
  • Every admin account, and every account an AI tool connects to, has MFA on.
  • You turned off at least one connection, and confirmed the AI can no longer reach it.

---

Next: the AI tools that aren't on your list yet, because nobody mentioned them.

→ Lesson 5: Find the AI tools nobody told you about

Previous: What your AI apps can see on each computer · Next: Find the AI tools nobody told you about