The Hacker NewsWednesday · July 1, 2026FREE

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

Microsoft has issued a warning about a novel attack vector targeting AI agents that use the Model Context Protocol (MCP). According to the advisory, attackers can poison MCP tool descriptions by embedding malicious instructions within the text that describes a tool's function. When an AI agent processes these descriptions, it may be tricked into performing unintended actions, such as leaking sensitive data to an external server. The attack exploits the trust AI agents place in tool descriptions, which are often sourced from untrusted third parties or user inputs. Microsoft recommends that developers rigorously validate and sanitize all tool descriptions before they are consumed by AI agents, and implement strict access controls to limit the potential damage. The company also suggests using least-privilege principles for agent permissions and monitoring agent behavior for anomalies. This advisory comes as the use of AI agents in enterprise environments grows, raising concerns about security and data privacy.