A field guide for people building with AI
A practical guide to AI guardrails
Understand what can go wrong. Put checks where they can stop it. Test what gets through.
Requested action
restart_service("api-prod")An example policy, not a live infrastructure request
What a guardrail does
A guardrail checks a request, a response or a proposed action against a rule. Some checks catch unsafe content. Others limit what an agent can do. The useful question is where the check runs and what happens when it fails.
Start reading
Build checks you can explain
What are AI guardrails?
Learn where AI guardrails check content, data and actions, and how to choose controls for a real application.
Prompt injection prevention for apps and agents
Reduce prompt injection risk by separating untrusted content from authority and testing the data and actions an attacker could reach.
Tool calling guardrails that hold at execution time
Validate tool arguments, check trusted permissions and bind approvals before executing an AI agent's proposed operation.
How to test AI guardrails
Build a repeatable guardrail test suite that measures prohibited outcomes, valid-task failures and behavior during outages.
Try the policy lab
Change the request. See the decision.
Choose an action and environment, then change the access and approval conditions. See which rule stops a request and why.
Try a policy checkPublished by the team behind Emisar, Ryker and co:op. The guides link to their sources and explain where each approach stops. How we write and review.