Researchers Trick Microsoft Copilot into Revealing the Secret Input That Enabled Its Own Exploitation

Security researchers at Varonis developed an exploit for Microsoft 365 Copilot that could exfiltrate sensitive user data after a victim merely clicked a link. Instead of using traditional vulnerability-hunting methods, they got Copilot to disclose the details of its own guardrails through a series of questions.

The AI assistant ultimately revealed an undocumented prompt parameter that completely bypassed the requirement for explicit user consent. The discovery highlights an unusual attack surface in which the AI itself becomes a source of critical security secrets.