The Emerged Security and Privacy of LLM Agent: A Survey with Case Studies

What is it about?

LLM agents are AI systems that plan, use tools, and interact with people to complete real-world tasks. As they spread into areas like education, coding, finance and customer service, new security and privacy risks are emerging. This survey maps the full threat landscape. We first explain how LLM agents are built and operate, then divide risks into two groups: threats inherited from LLMs (e.g., jailbreaking, data extraction, hallucination) and agent-specific attacks that target perception–thinking–action workflows (e.g., knowledge poisoning, output manipulation, and functional manipulation through tools and APIs). We analyze the real-world impact on people, the environment, and other agents, review existing defense strategies, and point to future trends. To make the topic practical, we include case studies that show how failures happen and how to mitigate them.

Featured Image

Photo by Growtika on Unsplash

Photo by Growtika on Unsplash

Why is it important?

Organizations are rapidly deploying LLM agents to automate complex tasks, but benchmarks show today’s agents still fail safety checks and can be coerced into harmful behaviors. Because agents can call tools, remember information and act over time, small prompts or poisoned knowledge can cause cascading, real-world consequences ranging from data leaks to unsafe actions. Our survey gives a concise map of risks and defenses so teams can assess exposure early, prioritize safeguards, and build more trustworthy LLM-agent systems.

Perspectives