MCP Security - The Exploit Playbook and How to Stop Them

Explore the critical security vulnerabilities in Model Control Protocol (MCP) implementations through this 27-minute conference talk that reveals how attackers exploit AI agent interactions. Learn about the alarming statistics showing 10% of the 13,000+ MCP servers launched in 2025 are malicious, while the remaining 90% contain exploitable vulnerabilities. Discover real-world attack vectors including tool poisoning where malicious instructions hide in tool descriptions, rug-pull tactics that change behavior after approval, conversation history theft, and cross-server tool shadowing techniques. Examine practical defensive strategies and implementation patterns to protect MCP deployments, whether you're securing AI agents as a security engineer, developing MCP servers, or integrating business systems like CRM with Claude. Gain comprehensive understanding of the MCP attack surface, practical knowledge of exploit mechanisms, a security checklist for implementations, and strategies for detecting and responding to MCP-based attacks in enterprise environments where adoption often outpaces security assessment.