Large language models (LLMs) are revolutionizing how we interact with technology. But what happens when the input—the prompt—gets lost? This is a critical challenge, especially with the rise of inference-only APIs that limit access to the inner workings of LLMs. Researchers have been grappling with prompt recovery, a task crucial for understanding LLM behavior, protecting privacy, and ensuring responsible use. Existing methods, like "jailbreak prompts," have proven unreliable. Now, a groundbreaking new technique called DORY (Deliberative PrOmpt RecoverY) offers a more effective solution. DORY leverages a surprising factor: the uncertainty of the LLM's output. By analyzing the probabilities assigned to different words in the output, DORY identifies clues about the original prompt. It reconstructs drafts, refines them with hints based on uncertainty, and filters out noise. This clever approach has achieved state-of-the-art results, outperforming existing methods by a significant margin. DORY's success opens exciting possibilities for enhancing LLM transparency and security. It could help prevent privacy leaks, identify harmful content generation, and even recover lost prompts for everyday users. While DORY has shown remarkable promise on current LLMs, its effectiveness on more advanced models like GPT-4 remains to be seen. Further research is needed to explore DORY's potential in the ever-evolving landscape of AI. But for now, DORY represents a significant leap forward in our ability to understand and control these powerful language models.