Microsoft Copilot revealed vulnerability through questioning
Researchers discovered a critical security flaw in Microsoft 365 Copilot Enterprise by asking the AI model questions.
Security firm Varonis has identified a critical vulnerability in Microsoft 365 Copilot Enterprise. Attackers could exploit this flaw to make the AI model reveal sensitive user data, including passwords, without explicit user confirmation.
The discovery method was unconventional; researchers directly questioned Copilot itself. Through a series of prompts, the AI divulged information about its safety mechanisms, ultimately leading to the revelation of an undocumented parameter that bypassed consent requirements.
Varos researchers aimed to create an exploit where a user could inadvertently expose data simply by clicking a link. While Copilot typically refuses commands without user consent, the researchers gathered intelligence on its guardrails and limitations through persistent questioning.
The identified parameter allows Copilot to execute commands and exfiltrate data without requiring the user to provide explicit confirmation. Microsoft is expected to address this security issue.