We introduce the "Deadlock Attack," which uses an adversarial embedding to force LLMs into perpetual thinking loops, leading to resource exhaustion.