Making Harmful Behaviors Unlearnable for Large Language Models