Researchers Reconstruct AI System Prompts From Model Output Alone

Researchers from IIT Bombay and Adobe Research have developed a technique called Previous-Token Prediction (PTP) that can reconstruct the original prompt used to generate an AI model's response, using only that response text. The method trains a reverse language model to predict preceding tokens rather than the next one, effectively reversing the standard text-generation process. Notably, the inverse model — built on the lightweight Qwen-3-0.6B — was able to recover prompts sent to GPT-4o without any knowledge of which model produced the output. The demonstration is currently limited to short, one-to-two sentence prompts, and the researchers have not tested it against longer system prompts or live commercial deployments. The findings raise concerns for businesses that treat system prompts as proprietary assets, as the research suggests that generated text alone may be sufficient to expose the underlying instructions.