💨 Abstract

Google is introducing agentic features in Chrome that automate tasks like booking tickets or shopping. To ensure security, Google uses several models: a User Alignment Critic to review tasks, Agent Origin Sets to restrict data access, and an observer model to check URLs. The system also seeks user consent for sensitive actions and employs a prompt-injection classifier to prevent unauthorized actions. Additionally, Google is testing these capabilities against potential attacks. Other AI browser makers, like Perplexity, are also focusing on security measures.

Courtesy: Ivan Mehta