AI Glossary Term
What does “Guardrails” mean?
Rules, checks, and controls that limit unsafe or undesired model behavior.
How Guardrails works
Example
A bank's assistant is set up so it will not reveal other customers' data and will refuse to give legal advice. One check screens the request before it reaches the model, and another screens the draft reply before the customer sees it.
Why it matters
Models can be steered into harmful or off-topic output. Guardrails add layers of control around them, though none is perfect.
Common questions
Where do guardrails sit?
Before the model, as input checks; inside it, via training and system prompts; and after it, as output filters.
Can guardrails be bypassed?
Sometimes. Jailbreak and prompt-injection attacks try to get around them, which is why layered defenses and testing are used.
Do guardrails make a model less useful?
They can, if too strict, by refusing legitimate requests. Teams tune them to balance safety and helpfulness.