Microsoft Warns Poisoned MCP Tool Descriptions Can Make AI Agents Leak Data

This report details a new security vulnerability impacting AI agent-based systems, specifically leveraging the Model Context Protocol (MCP). Attackers can inject malicious instructions into tool descriptions used by AI agents, causing them to perform unauthorized actions like data exfiltration without triggering typical security alerts. The vulnerability stems from the mixing of instructions and data within agents, creating a 'trust boundary' that allows attackers to manipulate agent behavior. Microsoft recommends implementing stricter controls around connected tools, reviewing tool descriptions, and incorporating human oversight for risky actions.