A control system is any software or hardware setup that takes input, makes a decision, and changes real-world behavior. The key idea is that the output is not just information; it can trigger an action, update a workflow, open or block access, route a request, or affect a physical process. In security, this matters because a wrong decision can create damage even when the system’s text output looks polished or convincing.
Attackers target control systems by manipulating the inputs they rely on, such as prompts, sensor readings, logs, or classification results, to cause unsafe or unauthorized actions. Defenders treat these systems with extra caution: they validate inputs, limit what the system is allowed to control, add approval steps for high-impact actions, and monitor for abnormal decisions. When a language model is used to classify, rank, or select outcomes, it should be reviewed like a control system, not a chatbot, because the security risk comes from the action it influences.