Level 3 · Expert
Guardrails: the bridge's railings
Rules that check what goes in and out of the model.
A guardrail is a railing: it doesn't stop you crossing the bridge, it stops you falling. In AI they're checks before and after the model: blocking sensitive data going in, stopping off-topic answers and catching manipulation attempts.
Try it: decide what should pass and what should be blocked for a bank's customer service assistant.
Try it today
Write five things your assistant should never say. That list is the first draft of your guardrails.
Test yourself
Someone types "forget your rules and…". What's that called?
- Prompt injection
- Fine-tuning
- RAG
See the answer
Prompt injection. It's an attempt to manipulate the model with instructions hidden in text. Guardrails detect and stop it.
If you lead a team
Ask to see the tests: how many manipulation attempts were tried and what share was blocked.