Approvals before sensitive actions
Actions are grouped by category: emails, sending, deleting, payments, system settings and routines. For each Bot, you decide which categories need your OK and which are free. Payments always need it.
An approval request shows what will happen: the recipient, the content, the amount. You approve or decline from the web, desktop or mobile apps. See how approvals work.
A second model reviews sensitive actions
Before a send, a deletion, a payment or a system authorization goes out, a second model reads the action next to your request. It looks for a mismatch (recipient, amount, content), for instructions that came from an email or a web page rather than from you (prompt injection), for a data leak or for a disproportionate action.
It never decides for you. Its verdict is shown on the approval request, and if it says “block”, an action that was free goes back to needing your approval.
Outside triggers are restricted
When work starts because of an incoming email or a webhook, the Bot runs in a restricted mode: it cannot send anything without your approval. That limits what a malicious email can make it do.
Passwords stay in a vault
Credentials you give your Bot are stored encrypted in a vault and filled in on its computer. The AI model never sees them, and they are never shown back in the conversation.
Your computer, your log
Each person has their own computer; it is never shared with another person. Every change in your account (Bots, approvals, vault, routines, devices) and every sign-in is recorded in an activity log you can read. The log holds labels and identifiers, never message contents or passwords.
What stays in your hands
An approval only protects you if you read it. Keep sensitive categories under approval while you get to know your Bots, and check the beta limits.