A field guide to AI guardrails.

What you can find here

A useful guardrail has a specific job: reject an unsafe request, check a tool argument, limit access to a resource or catch a bad result. The guides explain where each check belongs and how to test the cases it might miss.

Start with what AI guardrails do, then work through prompt injection prevention, tool permissions or guardrail testing. The policy lab lets you inspect a small decision rule, and the resource library provides worksheets and examples to adapt.

Use the examples in your own system

Each guide connects a failure to a control and a way to check it. Review the assumptions before copying an example: the same read operation may expose public status in one system and private records in another. Successful tests apply to the configuration and cases they covered.

Sources and review dates

Technical claims link to primary sources: standards, project documentation or the code that implements a behavior. Each guide has a review date, a source list and a downloadable Markdown version. Definitions and examples can be cited using the citation at the end of the article.

A review date records a content check. Product trials and deployment tests need their own evidence. Examples say whether they are conceptual, runnable or browser-only demonstrations; fictional records are labeled.

Product examples and ownership

A product belongs in a guide when its documented behavior helps explain the topic. Comparisons distinguish what the documentation says from what a reader still needs to test. Instructions and templates should remain useful with other implementations.

This publication shares ownership with Emisar, Ryker and co:op. Relevant examples disclose that relationship beside the mention.

Privacy

The interactive tools run in your browser. They do not send the values you enter to us. We do not add analytics, advertising scripts or tracking cookies. Cloudflare Pages serves the site and may process request data under Cloudflare's privacy policy.