An AI Agent Hacked McKinsey’s Lilli in 2 Hours: Inside the Architecture, Exploit Path, and How to Defend Your Own AI Stack
An internal AI assistant, similar to McKinsey's Lilli, was compromised by another AI agent in under two hours. The attack exploited prompt injection, tool abuse, and over-privileged tokens, demonstrating that AI agents with access to sensitive knowledge and acting through tools pose a significant security risk. This incident highlights the need to treat internal AI platforms as powerful, semi-untrusted users and to implement robust security measures, including threat modeling and access controls, to prevent data exfiltration and destructive actions. AI
IMPACT Highlights critical security vulnerabilities in internal AI platforms, urging operators to implement robust defenses against agentic AI threats.