Written by: Mio Kanzaki(Security Analyst (AI))Reviewed by: 千葉圭介Updated
Excessive agency is a state in which the permissions for operations given to an AI are too large. OWASP explains it as a weakness in which a system using an LLM (an AI that converses in text) performs a harmful operation after receiving unexpected, ambiguous or manipulated output.
How does excessive agency work?
It is like handing a new employee the key to the company safe and the authority to place orders with every business partner from the first day. Even without ill intent on their part, a misunderstanding or a clever lead from outside can cause an operation that cannot be undone.
OWASP divides the causes into three. One is a state in which functions that are not needed can be used. Another is a state in which the AI holds more permissions than needed in the system it operates. The third is a state in which operations with a large impact can be executed without independent confirmation or approval. Whatever caused the AI to malfunction, the size of the damage is decided by the systems the AI is connected to.
The AI Guidelines for Business (Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry) describe an AI agent as an AI system that senses its environment and acts autonomously toward a goal. The wider the range you hand over, the heavier this problem becomes. The organization of the topic is on the excessive agency vulnerability page.
Why does excessive agency matter?
AI is increasingly entrusted with real operations such as sending email, registering bookings and deleting files. Because it can operate, the effect of a mistake is no longer limited to a typo on a screen.
The measures OWASP shows are to give the AI only the functions it needs, to narrow even those functions to a range that suits the purpose, to keep permissions at the target of the operation to a minimum, and to put human approval in front of operations with a large impact. Checking permissions in the system that receives the operation is also included.
This idea is the same as the principle of least privilege. Give an AI, as with a human staff member, only the range the work needs. The behavior of an AI can also be taken over through prompt injection, which makes narrowing especially important. The free URL check shows what is visible from outside.
For a review of your setup, contact us.
Related terms
- AI agent: an AI that decides for itself and carries out work toward a goal.
- Least privilege: the idea of giving people and programs only the permissions in the range they need.
Next step
List the functions and permissions you have connected to the AI. Remove the functions you do not use, and put human approval on operations with a large impact. The organization is on the excessive agency vulnerability page, and what is visible from outside is covered by the free URL check. For a review, use a free consultation.
Frequently asked questions
Is narrowing permissions enough?
No. OWASP shows three together: hand over only the needed functions, keep permissions to a minimum, and put human approval in front of operations with a large impact. It also says not to leave this to the AI's own judgment, and to check permissions in the system that receives the operation as well.
If the AI does not make mistakes, is there no problem?
It can happen whatever the cause of the mistake. OWASP explains that whether the AI malfunctioned because of nonsense output or because of corrupted instructions, the size of the danger is decided by the systems the AI can touch.
Frequently asked questions
Is narrowing permissions enough?
No. OWASP shows three together: hand over only the needed functions, keep permissions to a minimum, and put human approval in front of operations with a large impact. It also says not to leave this to the AI's own judgment, and to check permissions in the system that receives the operation as well.
If the AI does not make mistakes, is there no problem?
It can happen whatever the cause of the mistake. OWASP explains that whether the AI malfunctioned because of nonsense output or because of corrupted instructions, the size of the danger is decided by the systems the AI can touch.