Research: Conventional Cybersecurity Won't Protect Your AI

"In June 2025, researchers uncovered a vulnerability that exposed sensitive Microsoft 365 Copilot data without any user interaction. Unlike conventional breaches that hinge on phishing or user error, this exploit, now known as EchoLeak, bypassed human behavior entirely, silently extracting confidential information by manipulating how Copilot interacts with user data. The incident highlights a sobering reality: Today's security models, which are designed for predictable software systems and application-layer defenses, are ill-equipped to handle the dynamic, interconnected nature of AI infrastructure."

"The incident highlights a sobering reality: Today's security models, which are designed for predictable software systems and application-layer defenses, are ill-equipped to handle the dynamic, interconnected nature of AI infrastructure."

In June 2025, a vulnerability exposed sensitive Microsoft 365 Copilot data without any user interaction. The exploit, named EchoLeak, bypassed human behavior and did not rely on phishing or user error. EchoLeak silently extracted confidential information by manipulating how Copilot accesses and processes user data. The event demonstrates that AI components can create novel attack vectors independent of human actions. Current security models prioritize predictable software behavior and application-layer defenses. Those models are insufficient for the dynamic, interconnected architectures of modern AI infrastructure, creating increased risk for sensitive data.

Read at Harvard Business Review