An agentic browser is a web browser with built-in AI that can read page content and perform actions such as clicking, typing, filling forms, and navigating on a user’s behalf. Unlike a normal chatbot, it operates inside the browser session, so it may inherit the user’s logged-in state, cookies, and access to sensitive sites.
This matters because the browser becomes an active security boundary, not just a display layer. If malicious page content uses prompt injection or indirect prompt injection, the agent can be persuaded to follow attacker instructions instead of the user’s intent. In real attacks, that can lead to credential theft, unauthorized form submission, or data exposure from authenticated accounts. Defenses focus on reducing agent power: use logged-out or separate profiles for risky tasks, require approval before sensitive actions, block dangerous domains, pause automation on finance or admin pages, and apply least privilege wherever possible.