AI SecurityPoisoned Tool Descriptions Turn Helpful AI Agents Into Quiet Exfiltration Channels
Microsoft Incident Response shows how a single malicious MCP tool description can coax an agent into leaking corporate data without tripping a single policy check.
