Semantic caching can reduce AI inference costs, but isolation failures can expose sensitive data and enable malicious responses or tool calls. NVISO explores these risks through proof-of-concept attacks and CVE-2026-89032, a cache isolation vulnerability discovered in LiteLLM.