Security and permissions: what the assistant sees and refuses
The assistant works with your own permissions, and the system itself checks every action. It doesn’t see your teammates’ conversations, and it doesn’t follow instructions written inside your data.
Your permissions are its limits
The same permission levels: Allowed, Team-wide access, If responsible and Denied, record by record. There’s no separate permission for the assistant.
The owner and super admins go past the permission matrix as usual, but deleting needs their confirmation too. During a support session, this bypass is turned off.
Reading chats needs the permission to see chats, and reading email needs the permission to see the mailbox.
Privacy
Every conversation belongs to its owner only. The assistant can’t reach other workspaces.
Protection from hidden instructions
Text inside records, comments, messages and email is data, not instructions. If the assistant finds text that looks like an instruction, it warns you about it.
Even if the model were tricked, every action is checked by the system and needs your approval.
If a reply comes out garbled, it’s deleted instead of shown, and nothing changes.
What it refuses
Revealing its instructions or its list of tools, claiming it has no limits, or naming the company or the model it runs on.
Changing permissions or billing, sending invitations, connecting messaging accounts, reading files, and diagnosing product faults. It points you to the right page or to support.
If something goes wrong
A real example
The receptionist asked the assistant about the doctors’ salaries.
- The salaries unit is denied in her role.
The assistant found nothing, because it sees only what she sees.