How AegisAI stops attacks on AI assistants
Episode details and transcript
A 48 second animated walkthrough of a prompt injection attack: an email carrying hidden instructions aimed at the AI assistants that read a mailbox. It shows the four steps AegisAI takes: Connect, Investigate, Protect, Extend. Illustrative example; names and details are fictional.
"Meeting notes from Tuesday" arrives with hidden white text that tells your AI mail assistant to forward contracts outside. A person reading it sees nothing wrong. Here is how AegisAI catches it.
01 Connect
AegisAI connects to Microsoft 365 or Google Workspace by API, out of band.
02 Investigate
Authentication passes, but agents flag a partner domain on a new server, hidden instructions written for an AI, and a policy that agents may not forward contracts. The verdict: prompt injection. The email is removed from 40 mailboxes and assistants are told to ignore it.
03 Protect
Inbound email security stops attacks that pass SPF, DKIM and DMARC.
04 Extend
The email is added to the red-team set, so your AI assistants are tested against it nightly.
Illustrative example. Names and details are fictional.
