活动预告｜CodeWisdom可信AI系统系列学术报告第14期：可信大语言模型智能体

Trustworthy LLM Agents: Uncovering and Securing the Attack Surface from a Systems Perspective

The rapid evolution of LLM foundation models and agents is reshaping daily life—from general-purpose AI assistants such as ChatGPT to specialized coding agents and autonomous deep-research agents. Yet this leap in capability introduces a new class of security and privacy challenges that traditional defenses were never designed to handle. Ensuring trustworthiness in LLM agents has thus become a fundamental and urgent problem for both academia and industry.

In this talk, I will present a systematic security framework for trustworthy LLM agents, organized around three critical and previously under-examined system modules that every production agent relies upon: (1) Prompt Compression, where we show that widely adopted compression modules (e.g., LLMLingua) expose a new attack surface that causes an adversarial information loss to manipulate LLM agent behavior; (2) Semantic Caching, where we reveal an inherent security vs. performance trade-off in the semantic cache design and demonstrate practical cache collision attacks that can hijack LLM responses and agent actions; and (3) Agentic Tool Invocation, where we show that a novel mode-behavior gap in the tool-invocation pipeline of major real-world coding agents(Cursor, Claude Code) can be exploited into security-sensitive system prompt leakage and a further Remote Code Execution(RCE). We also demonstrate the first query-agnostic indirect prompt injection that fires under arbitrary user queries.