What is LLM security?

LLM security is the practice of protecting AI systems built on language models from misuse, manipulation, and data leaks. Because LLMs behave differently from traditional software — they follow natural-language instructions and can take actions — they open up risks that ordinary security tools weren't designed for.

Why is securing an LLM different from securing normal software?

Traditional software follows fixed code paths, so you can lock down exactly what it will and won't do. An LLM doesn't work that way — it reads instructions written in plain English and decides how to respond, which means the "attack surface" is language itself. Anyone who can get text in front of the model, whether through a chat box, an uploaded document, or a webpage it reads, can potentially influence its behavior.

That blurring of the line between data and commands is the root of why LLM security needs its own playbook.

What are the main LLM security threats?

A handful of risks come up again and again:

• Prompt injection — hidden or malicious instructions, often planted in a webpage, document, or user input, that hijack the model into ignoring its rules or doing something harmful. This is the signature threat.
• Jailbreaking — tricking the model past its safety guardrails so it produces content it's supposed to refuse.
• Data leakage — the model revealing sensitive training data, system prompts, or another user's information.
• Agent misuse — an AI agent with real tool access being tricked into taking damaging actions, like deleting files or sending data out.

How do you defend an LLM system?

Because the model's own safety training isn't enough on its own, defense takes layers: validating and filtering inputs, checking outputs before they're shown or acted on, strictly limiting what tools and data an agent can touch, and monitoring for abuse. The guiding principle is to never fully trust the model's behavior, especially when it's handling untrusted input or wielding real-world capabilities. As AI moves from chatting to acting, LLM security is becoming as essential as traditional application security.