Wiki Home
LLM tool use
Definition
Tool use lets a language model act beyond producing text: it emits a structured call to an external function, such as a web search, an HTTP request, a code interpreter, an email send, or a key-value memory, and conditions its next generation on the returned output. The model decides which tool to call, when, and with what arguments, either through a prompting pattern that interleaves reasoning and actions or through learned tool-calling. Chaining tool calls toward a goal with little human oversight is what is usually meant by an LLM “agent.”
Two properties matter for security. Tool outputs re-enter the context as text the model tends to treat as trustworthy, and tool calls can produce side effects in the world, sending data, executing code, or writing to storage that later sessions read.
Papers that use this concept
- Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection — tools and memory turn a steered model into data exfiltration, persistence, and worming.