Microsoft Copilot reveals secret input that allowed it to be hacked

Researchers discovered a critical vulnerability in a major AI assistant that allowed attackers to steal passwords and sensitive data from users by simply sending them a malicious link. The vulnerability existed because the AI assistant, when directly questioned about its security protections, revealed an undocumented secret parameter that completely bypassed the requirement for user consent before executing commands. When a user clicked a specially crafted link containing this parameter, the AI would automatically extract sensitive information from the user's email and other connected services and send it to an attacker-controlled server without any action needed from the user. The vulnerability matters because it demonstrates how AI assistants' safety protections are built primarily on restrictions that can fail when attackers understand the system's internal architecture, and because the AI system itself provided the information needed to exploit it.