Is it safe to give an AI access to my email?

Does my AI see my email password?

No. Every connection happens on a page you open yourself: Google's or Microsoft's sign-in, or a private form for iCloud and IMAP. The password or sign-in goes straight into encrypted storage. It never passes through your AI.

What needs my approval?

• Sending is off by default. When it is on, each email waits for your approval, unless you turn on Send on my behalf in Settings. That lets an agent you pick send replies to people you have already emailed, after a hold time you set. You can cancel a held send, and you get an alert about it if you turned alerts on.
• Unsubscribing waits for your approval.
• Reporting mail as spam to your provider waits for your approval.
• Payments happen only on our website, by you.

A yes in chat does not count. You approve on the approval page or in the app with your password (or by signing in with Google again, if you use Google sign-in) or an optional passkey, so an AI cannot approve for you. See How do I add a passkey or remove a trusted browser?

Can I undo what my AI did?

Archive, Trash, and spam can be undone for 7 days. Trash is never a permanent delete. VIP mail is left out of bulk actions.

Can an email give my AI orders?

Mail is passed to your AI marked as untrusted content, with an instruction not to follow anything written inside. This makes it harder for a stranger's email to trick your AI. No guard like this is perfect, which is why sending, unsubscribing, and spam reports still need you.

How do I see and stop what my AI is doing?

• The Activity page lists connects, drafts, sends, and changes your AI made.
• Settings lists every AI app you connected with the MCP URL, each with Disconnect. API keys are under Advanced, each with Revoke.
• Disconnecting an inbox deletes its stored sign-in right away.

Does my mail train an AI model?

When your AI reads an email, that email goes to your AI provider (for example Anthropic or OpenAI) and is handled under their terms. Some let you choose whether your chats are used to improve their models. Check your provider's settings before you connect.

Inbox Triage never uses your mail to train or improve a language model or any AI app, ours or anyone else's, and does not run an AI model on your mail. The spam filter learns from your own spam and not-spam decisions and your Junk folder. It keeps counts of word pairs, not the messages. Those counts also feed a shared filter, except for Gmail, Google Workspace, Yahoo, and AOL mail, which only trains that inbox's own filter.

More detail: Security, Privacy Policy.