How Enterprise AI Platforms Get Hacked: Lessons from the McKinsey Lilli Breach

Enterprise AI platforms fail in predictable ways — and the McKinsey Lilli breach in February 2026 is the clearest case study yet of how a system deployed to 43,000 users can be fully compromised in under two hours through vulnerabilities that have been documented since the late 1990s. An autonomous security agent built by CodeWall extracted 46.5 million consulting conversations, 728,000 confidential documents, 57,000 user accounts, and — most critically — gained write access to all 95 system prompts controlling what the AI shows McKinsey's consultants ( source: codewall.ai ). McKinsey's standard scanner, OWASP ZAP, missed it entirely. This is the pattern that's going to repeat across enterprise AI deployments in 2026. ML engineers are building RAG systems without thinking like security engineers, and the attack surface is compounding faster than most teams realize. The Attack Chain That Took Down Lilli The entry point was embarrassingly simple. Of Lilli's 200+ API endpoints, 22 require