4 min read
Research A survey of query-efficient model extraction attacks against commercial LLM APIs — how adversaries can reconstruct a functional shadow model using only input-output pairs, the commercial and security risks this creates, and the defences providers are deploying.