> ## Documentation Index
> Fetch the complete documentation index at: https://docs.altrum.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Guardrails

> Understand what a guardrail checks, what it does when it fires, and how a draft becomes the live version.

A guardrail checks the text going into or out of an AI system: what people send it, what the model answers, or both. When it finds what it looks for, it fires and does one of three things:

* **Log only**: the text goes through, and Altrum records what it found.
* **Block**: the text is stopped.
* **Mask**: the sensitive part is replaced before the text goes on. Only the two data protection guardrails can mask.

Every AI system has the same eight guardrails, all off until you switch them on. See [Guardrail types](/guardrails/types.md).

## Draft and live

You change an AI system's guardrails in a draft, which enforces nothing. **Save draft** keeps your work. **Apply** turns the saved draft into the next numbered version and makes it live. A version never changes once applied; to go back, you roll back to an earlier one. See [Versions](/guardrails/versions.md).

The AI system's **Guardrails** tab always shows what is live, as **Live: version 3** or **Nothing is live**, and says **Changes saved, not applied** when the draft is ahead.

Check a draft with [Test guardrails](/guardrails/test.md) before you apply it.

## Where guardrails come from

* **From an assessment.** Submitting one switches on the guardrails its controls call for, in **Log only**, and the assistant suggests settings from your answers. See [Guardrail suggestions](/guardrails/suggestions.md).
* **By hand**, at any time, with or without an assessment. See [Configure guardrails](/guardrails/configure.md).

## Who can change them

Everyone in your organization can see an AI system's guardrails and versions. Owners, admins, Governance Leads and the AI system's owner can change, test, apply and roll them back.

## Whose guardrails apply

Each AI system has its own guardrails, and your organization has a set for traffic that belongs to no AI system. A key attached to an AI system is held to that AI system's guardrails only. See [Organization guardrails](/guardrails/organization.md).

## Guardrails and risk

Guardrails that block or mask cover the risks an assessment found; one in **Log only** covers them only partly. When every risk is covered, the AI system's rating after controls drops a level. Its **Risk** tab says which guardrail to change. See [Residual rating](/ai-systems/risk-rating.md#residual-rating).

## Where they are enforced

The AI Gateway enforces the live version on every request and reply. Applying a version sends it there straight away. See [Guardrails at runtime](/guardrails/at-runtime.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.