Security researchers at Varonis found a critical vulnerability in Microsoft 365 Copilot for enterprise that could leak passwords and other sensitive user data after a user clicked a link. Instead of using reverse engineering, the researchers extracted the exploit details by asking Copilot about its guardrails.
The model revealed an undocumented prompt parameter that completely bypassed the requirement for explicit user consent. The discovery shows an AI assistant divulging its own hidden safety controls through ordinary questioning.
Comments