Researchers Hack Microsoft 365 Copilot After AI Reveals Its Own Hidden Bypass

Security researchers at Varonis discovered a critical vulnerability in Microsoft 365 Copilot that allowed them to exfiltrate passwords and other sensitive data when a user merely clicked a link. Instead of using traditional hacking techniques, they extracted the information from Copilot itself by asking a series of questions about its safety guardrails.

The AI assistant disclosed an undocumented prompt parameter that bypassed the requirement for explicit user consent. This allowed the researchers to execute powerful commands automatically and steal user data.