PTP Achieves Black-Box LLM Prompt Reconstruction via Inverse Modeling
August 3, 2026
PTP introduces a functional approach to inverting black-box LLMs by training an explicit inverse language model from scratch using synthetic data from the target model. This method enables near-exact prompt reconstruction without requiring access to model weights or logits.
HOW THIS AFFECTS YOU
●
researcherThis method shifts prompt recovery from a semantic reconstruction task to a generative inverse-token prediction task.
●
policyThis technique highlights new risks for prompt leakage and intellectual property theft in black-box LLM APIs.