Approvals the model can’t skip
Paying for something, writing to someone new, deleting data, inviting someone and setting up anything that keeps running on its own always need your yes. These checks run in LVLZ’s code before each action, not by asking the AI to behave, so a prompt hidden in a web page or a file can’t talk its way past them. See Approvals and undo. Instructions that arrive from outside, like text in a forwarded message or a page your agent reads, are treated as information, not as orders from you.Secrets are sealed
- Passwords, payment cards, addresses and the keys for your connected services are encrypted at rest with AES-256-GCM.
- Each workspace or user has its own encryption key, so one leaked row can’t be opened with another owner’s key.
- When a browser task needs a password or a card, the browser on the server fills it in. The AI model sees that a field was filled, never the value.
- Keys for tools and payment providers are entered on one-time secure pages, never in the chat.
Your account
- You sign in with a code bound to your own phone number. Codes and access tokens are stored only as hashes.
- See and sign out of your sessions in the web app under Settings.
- Every read of workspace data is checked against the signed-in user. Your browser can’t write to the database directly; every change goes through LVLZ’s server, which checks who you are and what you may do.
Your data
- See, edit, delete and export what your agent remembers in the memory center.
- Your agent talks only with you and your workspace members, and never messages your clients on its own.
- Read the privacy policy and security page for more.