Google Chrome adds new security layer for Gemini AI agentic browsing

Google announced a new layered security architecture for Chrome’s Gemini-powered agentic browsing, featuring a separate ‘User Alignment Critic’ model to vet agent actions, ‘Origin Sets’ to restrict cross-site access, user oversight for sensitive steps, and a prompt-injection detection classifier. The approach targets threats like indirect prompt injection, phishing, and fraudulent actions, and is supported by automated red-teaming and rapid updates via Chrome’s auto-update. Google is also offering bug bounties up to $20,000 to encourage testing and hardening of this agentic browsing framework.