GitCloud Shield: phishing protection for mail and chats
Every incoming message from mail and chats goes through rules and a language model before a person sees it. The home DNS blocks dangerous links as well.
Screenshots of the panel running on sample data: made-up senders (.example domains, hidden numbers), made-up numbers, no database or network connection; missing Polish UI labels translated, the home protection part hidden.
Task
A message from a fake buyer of a listing (their own courier and payment before shipping) looked like an ordinary conversation. Mail filters do not see chats, and links from a chat pass through no filter at all.
Solution
The shield checks new messages from mail, chats and the agents' mailbox: rules in Polish, English and Russian (link shorteners, IP addresses, punycode, a brand on someone else's domain), domains from the CERT Polska, OpenPhish and Phishing.Database lists, and a language model when in doubt. ClamAV scans attachments before they reach the backup. A dangerous message never becomes an agent task, and the owner gets a warning without a quote. Fake domains go to the DNS block on the router, so a click at home or over the VPN opens nothing. The security centre collects it all with a score from 0 to 100 and ready advice on what to fix; a sender can be marked safe with one button.
Result
- Calibrated on 1057 real messages (631 from chats, 426 emails): one dangerous, the real one
- 0 false phishing list matches on 200 domains from a month of correspondence
- A warning before the message is opened, no message text in the database
My role
Design, rules and code of the shield, integration with mail, chats and the router, calibration on real correspondence.