AI-Based Browsers: Are They Really Safe?

Executive summary: AI-based browsers that embed LLMs create new structural security and privacy risks—especially agentic browsers that act autonomously—because untrusted content can be interpreted as intent and used to influence browser behavior. The report documents real-world issues (EchoLeak CVE-2025-32711, Perplexity Comet prompt injections, HashJack URL-fragment attacks, and Atlas omnibox confusion), explains how these enable data exfiltration and cross-tab access, and recommends controlled adoption, isolation of agentic browsers from sensitive sessions, treating AI outputs as untrusted, and improved visibility and governance.