LLM agents have been widely used in our daily lives through Vibe Coding, Vibe Research, Deep Research, OpenClaw, and Hermes Agent. The increasingly complex agentic tasks incur high token cost and inference latency. LLM caching mechanism is therefore introduced to provide an optimization phase so as to reduce redundant or repeated LLM inference on similar LLM tasks. Our recent work (published in ICML 2026, https://arxiv.org/pdf/2601.23088) shows that LLM semantic caching is inherently vulnerable to a cache collision attack, in which an adversary can craft a malicious query that unexpectedly collides with an arbitrary user's benign query, thereby hijacking their LLM response and agent behavior. However, to date, there is no practical defense against this attack despite the wide application of semantic caching. This project aims to explore practical and deployable defenses for cache collision attacks on the LLM semantic cache.