Most costly messages are not malicious. A password pasted into an email, a sharp reply written in a hurry, a price sent to the wrong customer. Atlas Guard looks at a message before it goes out and tells the sender, in a popup, while there is still time to fix it.
The company decides how strict it is
Guard has three modes. Off checks nothing. Coach shows a private warning and the sender decides; nothing is reported to anyone. Guard does the same, and a finding marked block must be fixed before the message can be sent.
What it checks
A few checks are built in: insults in English and Persian, and secrets such as passwords, API keys and card numbers (card numbers are validated, so an order number does not trigger it). A secret in an email to someone outside your domain is blocked. On top of that come your own rules: entries in the company knowledge that an admin confirmed and marked to be checked. A rule can list trigger words, which is fast and needs no AI, or it can be judged by a model from the rule's text.
Only confirmed rules count
Drafts, entries in review and retired entries are never used. Guard follows what the company wrote down, not the model's own opinion.
The audit mode, which would log events for admins with a notice to all users, is not built yet.