Operations Playbook

How do I check that an AI can only read one mailbox?

By Alexey Bulygin
An open green letterbox with a key beside a locked blue letterbox

Write down the permitted mailbox and actions before connecting the assistant. “Read only” should describe the granted access, not a promise in the conversation. Use harmless test mail to check that reading works, and inspect the authorization before trusting the same connection with customer messages.

You can give a compatible AI client read-only access in TrekMail, with no permission to send, edit, or delete messages. Access to email content requires Pro or Agency. The restriction belongs in the connection permissions, so check it during authorization as well as telling the assistant what you want in chat.

Define the data it needs

First decide which mailbox the assistant will use. If the task concerns one project, choose that project's address when your email is already organized that way.

Don't grant access to the whole account out of habit. It may contain other employees' mailboxes and other clients' domains. The scope should match the task you are giving the app.

Check what reading permission includes as well. Some permission options include additional mailbox data, such as contacts or calendars. Read the description before approval. “Read-only” describes permitted actions, but it doesn't necessarily mean access only to the text of one message.

Check the mailbox boundary and the action boundary separately

A connection can have the right kind of action but too much data access. For example, reading might be allowed across several mailboxes when your task concerns only one. It can also have the intended mailbox boundary but permit more actions than the task needs. “Read this one mailbox” therefore contains two requirements, and both belong in the authorization review.

Ask the person configuring access to explain them without technical shorthand: which mailbox can the client read, and which actions can it perform there? Keep the answer beside the connection type and scope names. That record is easier to assess than a screenshot containing a long list of permissions with no explanation of the task.

Do not substitute a visible tool list for this review. A client can display a command without having permission to complete it, while hiding a command does not by itself establish a server-side restriction. The permission check happens when the request reaches the service. What the assistant says it can do and what the service allows it to do are related questions, but they are not the same evidence.

Use the scope reference for the credential family actually involved. Infrastructure information about a mailbox is different from its message contents, and an account-key REST scope is not named identically to every message or OAuth scope. Keep the exact names in the configuration record, while describing their intended effect in ordinary language for the person approving access.

Allow reading without sending or changes

During connection, choose reading access and check that sending and message modification permissions are off. Deletion must also be prohibited for this task.

The exact permission set depends on the connection method. See the TrekMail permission reference. If a developer sets it up, ask them to list the available actions in ordinary language and identify the permitted mailboxes.

“Do not send anything” can be a useful working instruction. The technical restriction should exist independently of how the next user request is worded.

Compare the expected and observed result of each test

Prepare two harmless test mailboxes and record which one should be permitted. In the allowed mailbox, use a message with a recognizable subject and a short known sentence. A successful read should identify that actual message. This checks useful access rather than only proving that the client connected to something.

For the second mailbox, the expected outcome is refusal by the service. Ask the configurator to show the relevant request result if the chat interface only gives a general explanation. A statement such as “I will respect your privacy” may describe the assistant's intention, but it does not establish which mailbox boundary the server enforced.

Test the prohibited actions on expendable data as well, through the same connection. Record whether sending, modification, and deletion are rejected as intended. Do not promote the connection to customer correspondence merely because reading worked: that checks the permitted action and leaves the other half of the review unanswered.

If a result differs from the plan, stop using that connection for the real task and correct the authorization before repeating the test. After a permission change, check the current connection again rather than referring to an earlier successful screenshot. Keep the owner and revocation method with the results. Even a correctly restricted read connection passes the permitted content to its client, so revoking it later ends future access rather than retrieving text that has already been read.

Remember that reading also transfers data

Read-only mode prevents this connection from changing messages. The AI client still receives the content it can access in order to answer questions. Its own data processing terms need separate consideration.

Asking it to search messages from only one sender does not necessarily restrict server-side access to that sender. Do not assume limits that are absent from the actual settings.

The AI agent guide explains connecting and disconnecting apps. When a task ends or the person responsible changes, review access and revoke connections no longer needed.

Keep a small permission record

Use one line for each test: the mailbox requested, the action attempted, the expected result, and the service result you observed. Keep a refusal separate from an action the client could not attempt. The record should explain what was actually checked, not simply say that the assistant was configured safely.

  1. Record the connection type and scopes. For account-key REST message reads, the scope is mailboxes:messages:read, not mailboxes:read. The message/OAuth read scope is messages:read; omit sending and writing permissions.
  2. Check the approved mailbox boundary as well as the action. A tool hidden from the assistant is not the same as a server refusing it. An assistant saying “I cannot send” is not proof of enforcement.
  3. Check how to revoke this connection and note who owns it. Disconnecting later cannot retrieve message contents already passed to the AI service.

The permissions reference explains the credential families. Read access can still disclose sensitive correspondence; choose the mailbox and the AI service deliberately.

Related reading

Official references

Share this article

Sign in to TrekMail

Access your dashboard, mailboxes and DNS.

12 characters

Reset email sent

If an account exists for this email, we've sent password reset instructions.

By continuing, you agree to TrekMail's Terms and Privacy Policy.