Reducing the Scope of Language Models with Circuit Breakers