Google details security measures for Chrome’s agentic features (techcrunch.com)

Google has unveiled its comprehensive security framework for agentic features in Chrome, which aim to autonomously perform tasks like booking tickets and shopping. As these capabilities grow, so do concerns about user data security and potential financial risks. To address this, Google employs several protective measures, including a User Alignment Critic powered by the Gemini model, which evaluates planned actions and ensures they align with user goals. Instead of accessing full web content, the critic reviews only metadata, adding a layer of security against misuse. Significantly, Google introduces Agent Origin Sets to control the origins that agents can access, thereby limiting exposure to potentially harmful content and reducing the risk of cross-origin data leaks. Users are also placed at the forefront of the security model; for any sensitive operations, such as navigating to banking sites, agents will request explicit user permission before proceeding. Additional measures include a prompt-injection classifier to mitigate unwanted agent actions, reinforcing the importance of security as browsers advance towards more autonomous features. This proactive approach from Google sets a precedent for other AI-driven browsers, emphasizing the need for robust security in rapidly evolving technology.