ChatGPT Spills Secrets in Novel PoC Attack

Researchers from DeepMind, OpenAI, ETH Zurich, McGill and the University of Washington demonstrated a "top-down" attack that uses API queries to recover proprietary LLM architectural details (such as embedding projection matrices and hidden dimension sizes); they report recovering full projection matrices for some OpenAI models for under $20 and estimate other recoveries could cost under $2,000, raising concerns about model theft, reverse engineering, and reduced black‑box protections even though no active exploitation is reported.