AgentsAugust 12, 2026via The Decoder
Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy
Why it matters
A fundamental vulnerability in deployed LLM systems: adversaries can reverse-engineer proprietary prompts (jailbreaks, guardrails, agent instructions) from public outputs, threatening the security model that many enterprises depend on.
Key signals
- IIT Bombay and Adobe Research developed 'Previous-Token Prediction' method
- Reconstructs original prompts from LLM output with near-perfect accuracy
- Works without access to model weights
- Model-agnostic (works across different models)
- Implications for proprietary system prompts and agent instructions
- Published August 12, 2026
The hook
Researchers just cracked prompt recovery: proprietary system prompts can now be reconstructed from output alone with near-perfect accuracy—no model weights needed.
Researchers at IIT Bombay and Adobe Research have built an inverse language model that reconstructs the original prompt from an LLM's output with near-perfect accuracy. Their method, called "Previous-Token Prediction," doesn't need access to model weights and works across different models. For compa…