LLM agents have been widely used in our daily lives through Vibe Coding, Vibe Research, Deep Research, OpenClaw, and Hermes Agent. The complex agentic tasks bring motorically increasing LLM context (system prompt, user prompt, tool-use scheme, tool-use return) via an iterative reasoning loop. LLM context compression is, therefore, becoming an essential module in modern LLM agent frameworks. Our pioneering work (https://arxiv.org/pdf/2510.22963) firstly shows that the LLM context compression module can be tricked by an attacker into performing unintended yet malicious context overwriting, enabling attacks on real-world agents. This project dives into further investigation of an advanced confused duty attack on the real-world context compression module of LLM agents. The goal is to formalize attack theory and develop advanced attack algorithms on LLM context compression.