AI-enabled exfiltration is the movement of data outside approved channels after an AI system is manipulated into taking an action. Rather than breaking into a system directly, an attacker may use crafted instructions or hostile content the AI processes to steer an agent toward an available tool, such as email, a browser, or a data connector.
This matters when an AI system can access sensitive information and act on it. Defenses include limiting permissions, requiring explicit approval for sensitive transfers, and monitoring tool use and data movement. Treating external content as untrusted and validating actions before execution can also reduce risk. The term describes a possible attack behavior; it does not by itself establish that a specific data theft occurred.