Written by: Mio Kanzaki(Security Analyst (AI))Reviewed by: 千葉圭介Updated
An AI agent is an AI that decides for itself and carries out work toward a goal. The AI Guidelines for Business (Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry) define it as an AI system that senses its environment and acts autonomously to achieve a specific goal.
How does an AI agent work?
It works much like a secretary who is asked, "Book my trip for next week," then checks which days are free, picks candidates and goes as far as making the booking. It does not only answer questions. It builds the steps itself and actually acts.
The same guidelines also show, for reference, the definition in the international standard ISO/IEC 22989:2022. It describes an automated entity that senses its environment, responds to it, and takes actions to achieve its own goals. A note adds a broader and more advanced concept: a goal-driven system in which several AI agents make decisions and act autonomously is called agentic AI.
Many AI agents use an LLM, an AI that converses in text, as their brain. The danger of the operations you hand over is gathered on the excessive agency vulnerability page.
Why does an AI agent matter?
When an AI that only answers makes a mistake, a person who reads the answer can notice it. An agent goes on to real operations such as sending email or making bookings, so the effect of a mistake or a takeover becomes real harm as it is.
OWASP organizes, as excessive agency, the weakness in which a system using an LLM performs a harmful operation after receiving unexpected or manipulated output. The causes are three: unnecessary functions, more permissions than needed, and autonomous execution without confirmation.
For that reason, give an agent only the functions and permissions that the work needs, and put human approval in front of operations with a large impact. The idea is the same as least privilege. The problem of permissions is also explained on the excessive agency glossary page. The free URL check shows what is visible from outside.
To get help with how to proceed when introducing an AI agent, contact us.
Related terms
- Excessive agency: a state in which the permissions for operations handed to an AI are too large.
- Large language model (LLM): an AI that learns from large amounts of text and can converse and write.
Next step
Before introducing an AI agent, list the operations the AI can perform. Then decide which operations need human approval. The organization of the dangers is on the excessive agency vulnerability page. The state of your own site can be seen with the free URL check, and you can talk through how to proceed with introduction in a free consultation.
Frequently asked questions
How is an AI agent different from a chat-type AI?
The AI Guidelines for Business define an AI agent as an AI system that senses its environment and acts autonomously to achieve a goal. The difference is that it does not only answer questions but moves on its own toward a goal.
Does autonomous mean that people do not have to do anything?
No. The same guidelines note that autonomy includes not only highly autonomous states but also systems with some degree of autonomy. People decide how much to hand over.
Frequently asked questions
How is an AI agent different from a chat-type AI?
The AI Guidelines for Business define an AI agent as an AI system that senses its environment and acts autonomously to achieve a goal. The difference is that it does not only answer questions but moves on its own toward a goal.
Does autonomous mean that people do not have to do anything?
No. The same guidelines note that autonomy includes not only highly autonomous states but also systems with some degree of autonomy. People decide how much to hand over.