I'm not talking about HTTP headers.

The classifier is a model; it examines the actual prompt.

They already do this for the safety "guardrails".