Google is enhancing Chrome’s security with layered protections for its new agentic capabilities, aiming to prevent indirect prompt injections and data leaks. These defenses include a specialized AI model called User Alignment Critic and expanded origin isolation policies. #Gemini #AgenticBrowsers
Keypoints
- Google has introduced layered security protections for Chrome’s new agentic features.
- The User Alignment Critic AI model vet actions to prevent goal-hijacking and data exfiltration.
- Chrome now uses Agent Origin Sets to limit agent data access to relevant and user-shared origins.
- Asking user confirmation is required before sensitive actions like site navigation or payments.
- The system includes prompt-injection classifiers to detect malicious content and prevent harmful actions.