Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy

Researchers at IIT Bombay and Adobe Research have developed an inverse language model capable of reconstructing original prompts from LLM outputs with near-perfect accuracy. This method, known as "Previous-Token Prediction," operates without needing access to model weights, making it applicable across various models. This technique is particularly significant for companies relying on proprietary system prompts, as it poses a serious security risk. The ability to reverse-engineer prompts from outputs means sensitive information could be compromised, prompting companies to reassess their security strategies.