From human clicks to machine intent: Preparing the web for agentic AI

The emergence of agentic browsing signifies a fundamental shift in how AI-driven agents interact with the web, moving beyond passive page viewing to actively executing user intents through tools like Comet and Claude browser plugin. These agents can perform complex tasks such as content summarization, email drafting, and booking services, but current web architecture is ill-equipped to support their needs, exposing vulnerabilities in security and control. Experiments reveal significant risks associated with this paradigm, including agents executing hidden instructions embedded in web pages or emails without validation, leading to potential privacy breaches and malicious actions. For instance, hidden commands can prompt agents to