New research shows how encoder models can predict LLM output length to prevent malicious requests that overwhelm AI systems.