Home / Glossary / AI agent security
Glossary
What is AI agent security, and what are its risks?
AI agent security is the practice of finding the AI agents a company uses, limiting what each one can reach and setting rules for what it may do. An agent does not just answer. It takes actions on a person's behalf, so the risks are about reach and permission, not only about what it says.
By the AIBound team · Updated 15 November 2026
Key takeaways
- An agent takes actions on a person's behalf, so the risks are about reach and permission.
- An agent can use the same logins and files as its owner.
- AI agent security finds each agent, maps what it can reach and sets a rule for each action.
Why AI agent security matters
Employees now run coding agents, desktop assistants and personal agents on their own devices. An agent can use the same logins and files as its owner. If it is given too much reach, a wrong or manipulated instruction can touch real systems. Security needs to know which agents exist, what each can reach and who is responsible for it.
How AI agent security works
AI agent security follows the agent from the person to the systems it can touch. It asks which agents are running, what identity and connections each one uses, and what actions it is allowed to take.
- 1 Find every agent on employee devices, with the person who runs it.
- 2 Map what each agent is wired into and how far it could reach.
- 3 Set policy inside the agent, so each action is allowed, asked about or denied.
AI agent security vs LLM security
LLM security protects a model and the text going in and out of it. AI agent security protects what an agent can do once it has tools, logins and files.
| Question | AI agent security | LLM security |
|---|---|---|
| What is the focus? | Actions, reach and permissions | Prompts, outputs and model behaviour |
| What is the main worry? | An agent doing too much, or the wrong thing | A model saying something harmful or leaking data |
| Where does control happen? | Around the agent, on the device it runs on | Around the model, at the prompt and the response |
Sources
- OWASP LLM06: Excessive Agency · genai.owasp.org
Sources checked 19 November 2026
Examples
AI agent security examples
FAQ
Frequently asked questions
What is AI agent security?
AI agent security is how a company finds the AI agents it uses, limits what each one can reach and sets rules for what it may do. It focuses on actions and permissions, because an agent acts for a person rather than only answering questions. The work is done on the device where the agent runs.
What are the main AI agent security risks?
The main risks are agents with too much reach, risky skills that carry hidden instructions, and agents nobody knew about. OWASP calls the reach problem excessive agency: an agent given more functions or permissions than its task needs. Each risk grows when no one can see which agents exist.
How is AI agent security different from LLM security?
LLM security looks at the model, the prompt and the answer. AI agent security looks at what the agent does: which systems it reaches, which actions it takes and under which login. A safe answer from a model can still lead to an unsafe action by an agent.
How do you secure AI agents on employee devices?
Find every agent, map what each one can reach, then set policy for each action so it is allowed, asked about or denied. Review the skills an agent installs before they run. Doing this on the device covers personal and coding agents that never pass through a network gateway.