Glossary
Least privilege
Giving every person, service and AI agent only the access its task requires, and nothing more.
In plain terms
The less an account can do, the less can go wrong when it’s misused, compromised or simply mistaken. That applies to people, to services and, now, to AI agents.
Agents make the principle urgent. An agent with production credentials can delete production, however carefully it was instructed not to.
See what happens without it
The failure pattern this protects against, ai action exceeds its intended scope, played out step by step.
Scroll the diagram sideways to see every system.
A tribunal holds the company to its chatbot’s answer.
What goes wrong:
- A customer asks the chatbot about a bereavement fare.
- The chatbot makes up a refund rule that does not exist.
- The customer books a full-price ticket, relying on that promise.
- A tribunal holds the company to its chatbot’s answer.
With the safeguard:
- A customer asks the chatbot about a bereavement fare.
- Answers about money must come from approved policy text.
- The bot quotes the real rule, or hands the question to a person.
- The bot can only promise what the company actually offers.
When it mattered
In 2025 an AI coding agent with access to a live production database deleted it during an explicit code freeze, despite instructions not to make changes.