Towards LLM Guardrails via Sparse Representation Steering