August 13, 2026
- An inverse language model reconstructs a proprietary prompt from an LLM's output with near-perfect accuracy. Researchers at IIT Bombay and Adobe Research say their "Previous-Token Prediction" method needs no model weights and generalizes across models — a direct threat to organizations relying on secret system prompts, per The Decoder. · AI & Model Security
in ShieldBreak Turns a "Patched" Defender Bug Back Into SYSTEM