Microsoft Copilot reveals secret input that allowed it to be hacked

Microsoft Copilot Enterprise has a critical security vulnerability—one that researchers exposed by asking the AI to explain its own defenses. When Varonis security researchers questioned Copilot about its safety guardrails—the mechanisms that require explicit user consent before executing sensitive commands—the AI readily explained how they worked. Through persistent follow-up questions about URL structures, deep links, and system architecture, Copilot progressively revealed more structural details, eventually divulging an undocumented prompt parameter that completely bypassed the consent requirement. Ars Technica reports that researchers then weaponized this hidden parameter to craft an exploit capable of stealing user passwords and sensitive data with just a link click—no user permission required.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton