Researchers demonstrated that a prompt could instruct Microsoft Copilot to search a victim's inbox, extract sensitive data, and leak it to an attacker-controlled server