New Model Reconstructs LLM Prompts with Near-Perfect Accuracy
TL;DR. Researchers developed a method to reconstruct large language model prompts from output text with high accuracy, posing a security risk for proprietary systems. - The "Previous-Token Prediction" inverse model trains on synthetic data to predict prior tokens without accessing model weights. - It can reconstruct exact prompts and semantic variants, even across different LLM architectures like Qwen and GPT-4o. - This technique threatens proprietary system prompts, potentially exposing trade secrets and moderation rules.
- Researchers developed an inverse language model, 'Previous-Token Prediction' (PTP), that reconstructs original LLM prompts from output text.
- PTP works without access to model weights and can reconstruct prompts for different LLMs, including proprietary ones like GPT-4o.
- The method can generate the exact original prompt and multiple semantically similar variants from a single output.
- This capability creates a significant security vulnerability, as companies' proprietary system prompts, including trade secrets and moderation rules, could be exposed.