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

Microsoft researchers have discovered a method by which attackers can exploit AI agents to leak company data. This is achieved by poisoning tool descriptions, which causes the AI agent to quietly hand over sensitive information to an outsider. The agent's actions appear routine and do not break any rules, making it unlikely to trigger alarms in a default setup. The research was conducted by Microsoft Incident Response. The attack relies on manipulating the AI agent's behavior without violating its rules. This technique allows attackers to obtain company data without raising suspicion.