CachePrune: Multi-Tenant LLM Serving Has a Side-Channel Problem.

New research shows unrestricted cross-user KV cache sharing in LLM serving creates side-channel attacks where adversaries infer other users' inputs by probing for cache reuse. CachePrune introduces privacy-aware sharing that prevents leakage while preserving 85%+ cache reuse. If you're running multi-tenant inference, this paper is required reading.