Microsoft Copilot Tricked Into Revealing Hidden Input That Enabled Data-Theft Exploit

Researchers at security firm Varonis created an exploit that exfiltrates user data when a user simply clicks a link. After Copilot refused, citing the need for explicit user consent, they questioned it about its own guardrails.

Through this dialogue, Copilot disclosed an undocumented prompt parameter that bypassed the consent requirement. The discovery allowed the researchers to pull off the hack, demonstrating a novel way to expose critical AI security flaws.