Home / Glossary / LLM security
Glossary
What is LLM security?
LLM security is the practice of protecting large language models, and the apps built on them, from attack and misuse. It covers the model, the data it learns from, the prompts it receives and the output it returns. Most of this work is about the AI a company builds, which is a different job from securing the AI that employees run on their devices.
By the AIBound team · Updated 17 November 2026
Key takeaways
- LLM security protects the model, its data, the prompts it receives and the output it returns.
- The OWASP Top 10 for LLM Applications lists the best-known risks, and prompt injection is first on the list.
- LLM security covers the AI a company builds, so the AI employees install on their devices needs its own check.
Why LLM security matters
Companies now put large language models (LLMs) inside support chatbots, search tools and coding tools. An LLM takes its instructions in plain language, and it reads instructions and data in the same stream of text. A line hidden in a web page or a document can steer the model as if the owner had written it.
The OWASP Top 10 for LLM Applications names the risks that follow. The 2025 list starts with prompt injection and includes sensitive information disclosure, supply chain risk and excessive agency. The stakes rise when a model is given tools, because a model that can send email or run code can do real damage when it is misled.
How LLM security works
LLM security puts checks around the model at each point where something can go wrong. Teams then test the checks by attacking their own app, which is called red teaming.
- 1 List the models and apps in use, and note what data and tools each one can reach.
- 2 Check what goes in. Filter prompts, and treat text from documents, web pages and tools as untrusted.
- 3 Check what comes out before it is shown to a user, stored or run as code.
- 4 Limit what the model can do. Give it the fewest tools it needs, and ask a person before a risky action.
LLM security vs endpoint AI security
LLM security protects the models and apps a company builds or hosts. Endpoint AI security protects the company from the AI its employees run on their own laptops. Most companies need both, and the two are often owned by different teams.
| Question | LLM security | Endpoint AI security |
|---|---|---|
| What does it protect? | The models and apps a company builds or hosts | The company, from the AI agents, MCP servers, skills and models employees run |
| Who usually owns it? | The app security team and the people who build the app | The security team that looks after employee devices |
| Where do the controls sit? | In the app, around the model's input and output | On the device, inside the agent |
Sources
- OWASP Top 10 for LLM Applications · genai.owasp.org
Sources checked 17 November 2026
Examples
LLM security examples
FAQ
Frequently asked questions
What is LLM security?
LLM security is the practice of protecting large language models and the apps built on them. It covers the model itself, its training data, the prompts it receives and the output it returns. The aim is to stop attackers misleading the model, stealing data through it or using it to reach other systems.
What are the biggest LLM security risks?
The OWASP Top 10 for LLM Applications is the most used list. Its 2025 edition starts with prompt injection, where hidden text steers the model. It also covers sensitive information disclosure, supply chain risk, data and model poisoning, improper output handling and excessive agency, where a model has more power than it needs.
What is the difference between LLM security and AI security?
AI security is the wider term. It covers every kind of AI system, including older machine learning models, and the ways people use AI at work. LLM security is the part that deals with large language models and the apps built on them, such as chatbots, assistants and coding tools.
Does LLM security cover the AI tools employees use?
Usually only in part. LLM security programs focus on the apps a company builds. The coding agents, MCP servers, skills and local models that employees install on their own laptops sit outside those apps. AIBound finds that AI on the device and stops the high-risk ones in real time, inside the agent.