AI glossary
Guardrails
Checks and limits placed around an AI system to keep its inputs and outputs safe, on-topic and within policy.
Guardrails can be simple (a list of allowed actions, a spending cap, requiring human approval before sending an email) or use another model to screen inputs and outputs for harmful content, leaked secrets or off-topic requests.
For agents, the most effective guardrails are often permissions: give the agent only the tools and access it truly needs, and ask for confirmation before irreversible actions.
Example: A shopping agent can search for products and add them to the basket, but not pay: payment needs a person to confirm it, and there is a spending limit per order.
In practice
- Combine several layers: permissions, limits, automatic checks and human approval.
- Test your guardrails by trying to get around them, as an attacker would.
- Log what the agent does so you can review it later.
How to use agents safely: what an AI agent is.
