← AI Glossary

Least privilege (AI agents)

Sovereignty & security

The principle of least privilege holds that every program and every user should have only the rights strictly needed for their task. Formulated in 1975 by Jerome Saltzer and Michael Schroeder, it becomes central with AI agents, which can make mistakes or be manipulated through prompt injection: an agent that summarises emails has no need to be able to send or delete them. The measures OWASP recommends against excessive agency follow directly from it: this risk, which OWASP attributes to excessive functionality, excessive permissions or excessive autonomy, moved from sixth to third place in its Top 10 for LLM applications in the 2026 edition. In June 2026, Microsoft also urged organisations to aim for “least agency”: even an agent with reduced rights can cause damage if it acts without oversight.

In practice at Gensai

In its agent projects, Gensai defines with the client access rights that belong to the agent and are limited to its task: read-only when that is enough, a list of allowed actions, human validation of sensitive operations and a log of actions.

How should an AI agent's rights be set?

Starting from the task, tool by tool: read-only if the agent only needs to consult, access limited to the data of the person it acts for, human validation for anything irreversible, such as a payment, a message sent or a deletion. ANSSI recommends not letting an AI system automatically carry out critical actions on the information system, and reviewing the rights of generative AI tools regularly, for example every month. A useful rule of thumb: avoid giving a single agent access to private data, exposure to untrusted content and the ability to send information outside, a combination Simon Willison named the “lethal trifecta” in June 2025.

Common mistake

Connecting the agent with an administrator account or a shared API key “to make it work”, counting on its instructions to hold it back. An instruction can be bypassed, a missing permission cannot: it is the connected system that must refuse the action, not the model that must decide to refrain.

Definition written with AI assistance, checked by Sébastien Lecocq, co-founder of Gensai.

Going beyond the definition?

From concept to project: Gensai builds custom, sovereign, GDPR-compliant AI solutions.

Related services: Chatbots and answer engines, Custom AI solutions

55 boulevard de Strasbourg, 75010 Paris
Information used only to answer the request.