AI Agent Breached McKinsey's Internal AI Platform Lilli
An autonomous AI agent developed by security startup CodeWall breached McKinsey's internal AI platform Lilli in a controlled security test, gaining access to millions of employee chat messages and sensitive file metadata within two hours. McKinsey confirmed the vulnerability, patched it, and stated that client data and the actual sensitive files were never at risk.
What happened
CodeWall, a security startup that uses AI agents to test customer security, conducted an authorized security exercise where its autonomous AI agent targeted McKinsey's internal AI platform Lilli. The agent gained full read and write access to Lilli's production database within two hours, discovering API documentation that exposed over 200 system endpoints.
The breached agent accessed 46.5 million internal employee chat messages, a list of 728,000 sensitive file names, 57,000 user accounts, 384,000 AI assistants, and 94,000 workspaces. It also exposed Lilli's internal system prompts and AI model configurations, revealing how the system was instructed to behave and what guardrails were in place.
Reported impact
McKinsey confirmed the vulnerability and was alerted at the end of February. The firm stated that while sensitive file names were visible, the actual sensitive files were stored separately and were never at risk, and that no unauthorized third party accessed confidential information or client data.
Index analysis
CodeWall described the accessed data as the firm's 'intellectual crown jewels' and 'the full organisational structure of how the firm uses AI internally.' The incident demonstrates how autonomous AI agents could autonomously discover and exploit vulnerabilities in other AI systems within hours.
Governance controls
McKinsey patched the identified vulnerabilities after CodeWall's findings. The platform processes over 500,000 prompts monthly and is used by more than 70 percent of McKinsey's workforce, over 40,000 employees.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.