Chrome is introducing new security measures for its AI-powered agentic browsing capabilities. A primary threat is indirect prompt injection, where malicious sites can trick agents into harmful actions. To combat this, Chrome employs a layered defense strategy. A key innovation is the User Alignment Critic, a separate AI model that vets agent actions for task alignment, preventing unintended consequences. Chrome is also extending its origin-isolation principles with Agent Origin Sets. This restricts agents to interacting only with relevant websites and data for their current task. These origin sets are divided into read-only and read-writable categories, limiting data exposure. Furthermore, Chrome ensures user transparency and control through a detailed work log and user confirmations for sensitive actions. Deterministic checks and model-based classifiers identify and prevent prompt injection attempts. Continuous auditing and red-teaming are used to test these defenses against evolving threats. Chrome also collaborates with the security research community, offering rewards for identifying vulnerabilities. The goal is to provide a secure foundation for agentic AI experiences in Chrome.