The integration of AI features into web browsers are useful for users, but how about for attackers? In this talk, I will explain the security implications of integrating the AI-powered Edge Copilot into the Microsoft Edge browser, by showcasing multiple vulnerabilities I discovered. Such as stealing content of any site, accessing microphone and camera permissions without user’s permission, and so on. This talk also highlights the interesting integration of a relatively secure system (Edge) with an insecure system (Bing), and how a mitigation enforced by a secure system can be circumvented using an insecure system. Additionally, I will demonstrate how LLM-specific exfiltration techniques can be employed to bypass traditional security mitigations.