A request classifier is a safety layer that inspects a user prompt before the model answers and assigns it a risk label, such as benign, dual-use, or potentially harmful. It may look for intent, keywords, technical context, and patterns associated with misuse. The classifier does not generate the answer itself; it helps decide whether the request should be answered normally, limited, routed to a safer path, or refused.
In cybersecurity, request classifiers matter because the same model can be useful for defensive work and dangerous for offensive guidance. A prompt about code review may be allowed, while a prompt that asks for exploit instructions may be blocked. In real systems, this control helps reduce abuse, but it is not perfect: attackers may try to disguise malicious intent, and defenders need classifiers tuned well enough to avoid blocking legitimate research.