Microsoft Copilot Vulnerability Exposes User Data Without Consent

Researchers at security firm Varonis uncovered a critical flaw in Microsoft 365 Copilot Enterprise that lets attackers extract user passwords and other sensitive information without any user confirmation. Rather than reverse‑engineering the model, the team leveraged Copilot itself, asking a series of questions about its safety guardrails. The assistant, which normally requires explicit user gestures to execute powerful commands, revealed an undocumented prompt parameter that bypasses this requirement. By exploiting this parameter, an attacker could trigger Copilot to reveal data simply when a user clicks a link. The discovery highlights a serious security gap in the AI assistant’s design and underscores the need for Microsoft to patch the vulnerability promptly.