To the surprise of no one in the security industry, processing untrusted, unvalidated input is a bad idea. Until about a week ago, Perplexity's AI-based Comet browser did just that - asked to summarize a web page, the AI-powered browser would ingest the text on the page, no questions asked, and process it. And if the page text - visible or hidden - happened to include malicious instructions, Comet would attempt to comply, carrying out what's known as an indirect prompt injection attack.
"Attackers are no longer simply experimenting with open source. The numbers are telling us that threat actors have identified data as the most profitable target, and developers as the easiest way in."