A summary that goes off track
You ask AI to summarize a web page and it praises an unknown product. That may be prompt injection.
HOW
White text on a white page
Bad actors hide commands in web pages or documents that people can't see but AI can read.
DANGER
When AI has tools
It's most dangerous when AI can read your email or click on the web, because a hidden command may tell it to send data out.
DEFENSE
Unknown documents are read-only
With unfamiliar documents, tell AI to only summarize and not follow any instructions inside.
This article is based on the video Prompt injection from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.