Legitimate permissions are the access rights granted to an agent or software system for its intended work. An AI agent might be allowed to read email, open files, send messages, or use connected tools. These rights are legitimate because they were authorized; the risk is that the agent may use them in ways the user did not intend.
In an attack, hostile instructions hidden in an email, document, or webpage can manipulate an agent into using its existing access to disclose data, trigger workflows, or delete files. The attacker may not need to break into the system or obtain new privileges. Defenses include granting agents only the minimum permissions they need, treating external content as untrusted, logging actions, and requiring independent confirmation for sensitive operations.