Tool misuse is the abuse of an AI system’s authorized actions, such as reading files, sending messages, querying databases, or triggering workflows, to produce outcomes the operator did not intend. The model may be behaving “correctly” from a permissions point of view, but the result is still unsafe because the action was taken in the wrong context or for the wrong goal.
In cyber security, tool misuse matters because agentic systems can turn a prompt problem into an access problem. If an attacker can inject instructions, confuse the agent, or manipulate its inputs, they may steer the system into leaking data, changing records, approving requests, or running tasks outside policy. Defenders reduce this risk with least privilege, approval gates for sensitive actions, scoped tool access, strong logging, and clear boundaries on what each agent may do.