New method reconstructs an LLM's hidden system prompt from its output text alone

Researchers built a technique that reverse-engineers an AI's secret prompt just by reading what it says, no model access required.