A browser-capable agent is an AI assistant that can fetch webpages, follow links, or use web search as part of its task flow. Unlike a chat-only model, it can process untrusted web content and sometimes take actions based on what it reads. That extra capability makes it more useful for summarizing pages or finding information, but it also expands the attack surface.
In cyber security, the main risk is indirect prompt injection: a malicious page can hide instructions that steer the agent away from the user’s intent. If the agent also has access to session context, account details, or other private data, those instructions may cause leakage or unsafe tool use. Defenses include isolating fetched content from privileged context, separating read-only analysis from action-taking, limiting outbound actions, and logging tool calls for review. In practice, browser-capable agents must treat the web as hostile input, not as trusted reasoning material.