An LLM application is software that uses a large language model to process, generate, summarize, or classify text. It may act as a chatbot, a content scanner, a document reviewer, or part of an automated agent that can call tools, search files, or send responses.
In cyber security, LLM applications matter because they often read untrusted input. If attacker-controlled text is fed into the model, it can try to steer the model with prompt injection, causing misclassification, hidden instructions, or unsafe tool use. This becomes especially risky when the application connects the model to email, web access, APIs, or other privileged actions. Defenses include separating instructions from data, limiting tool permissions, requiring human approval for sensitive steps, and using deterministic checks for high-risk decisions.