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

# Guardrails

> Configure global guardrails and file upload security checks to apply across all chats and agents at once.

<Info>Admin › Settings › Guardrails</Info>

**Global guardrails** are safeguards that apply to **all chats and agents at once**, regardless of individual agent guardrail settings.

* Use them to enforce a minimum input/output validation baseline across the entire organization.

<Frame caption="Admin > Settings > Guardrails — global guardrails · file upload checks">
  <img src="https://mintcdn.com/cloocus/3CYTE_uN2N3A4joq/images/admin/settings-guardrails.png?fit=max&auto=format&n=3CYTE_uN2N3A4joq&q=85&s=6ca60c466e1d2d86f6a5914add0fe151" alt="Guardrails settings" width="2880" height="1800" data-path="images/admin/settings-guardrails.png" />
</Frame>

<Note>
  What this tab shows depends on your [license](/en/admin/settings/license) feature keys. The tab itself requires the `file_guardrail` module, and the **Global Guardrails** section requires `global_guardrail`.
</Note>

***

## Global Guardrails

| Setting               | Description                                                                                                        |
| --------------------- | ------------------------------------------------------------------------------------------------------------------ |
| **Global Guardrails** | Overall ON/OFF toggle. When on, applies to all chats and agents                                                    |
| **Chat Guardrail**    | Select one guardrail to apply globally. Its PII detection, blocked words, and LLM Judge are then enforced globally |

<Note>
  The guardrails themselves (the check rules) are defined in [Workspace › Guardrails](/en/workspace/guardrails). This page is where you choose which of them to **enforce globally**.
</Note>

***

## File Upload Guardrails

Apply security checks to uploaded files.

* Turning the toggle on reveals the **Applied Scopes** and the checks below; the checks run **before** a file is processed.

### Applied Scopes

Choose which upload paths the guardrails apply to — **Chat · Knowledge · Project**. Unchecked contexts skip all guardrail checks.

### Checks

| Check                       | Form             | Description                                                                                                                                 |
| --------------------------- | ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------- |
| **Text Content Guardrails** | Guardrail picker | Scans the extracted document text with the selected guardrail. If a violation is detected (PII, sensitive data, etc.) the upload is blocked |
| **EXIF Metadata Stripping** | Toggle           | Automatically strips GPS, camera info and other EXIF metadata from uploaded images                                                          |
| **Macro Detection**         | Toggle + action  | Detects executable macros embedded in documents                                                                                             |
| **NSFW Image Detection**    | Toggle + model   | Uses an LLM to judge inappropriate images                                                                                                   |
| **Document Classification** | Toggle + model   | Uses an LLM to classify document sensitivity and act per category                                                                           |

<Note>
  **Text Content Guardrails**, **NSFW Image Detection** and **Document Classification** require a model or guardrail to be selected. **EXIF Metadata Stripping** and **Macro Detection** are static checks that run without a model.
</Note>

### Macro Detection

When the toggle is on, an **Action** can be chosen.

| Action    | Result                           |
| --------- | -------------------------------- |
| **Block** | The upload is blocked            |
| **Flag**  | The upload is allowed but marked |

### NSFW Image Detection

| Setting            | Description                                       |
| ------------------ | ------------------------------------------------- |
| **Model**          | LLM used for the judgment                         |
| **Prompt**         | Prompt that defines the judgment criteria         |
| **Pass Examples**  | List of image descriptions that should pass       |
| **Block Examples** | List of image descriptions that should be blocked |

### Document Classification

| Setting                            | Description                                                               |
| ---------------------------------- | ------------------------------------------------------------------------- |
| **Model**                          | LLM used for classification                                               |
| **Max Characters**                 | Upper bound on the body text passed to the classifier                     |
| **Prompt**                         | Prompt that defines the classification criteria                           |
| **Categories**                     | Set `ID` · `description` · **action** (Allow / Flag / Block) per category |
| **Pass Examples · Block Examples** | `Sample text` + `category` pairs that reinforce the criteria              |

***

## Guardrail Application Priority

Guardrails can be set at multiple layers. Rather than only one applying by priority, **guardrails from all layers are combined (union) and applied together**.

```mermaid theme={null}
flowchart LR
    A[Agent guardrails] --> R([All layers<br/>combined])
    G[Group guardrails] --> R
    O[Organizational unit guardrails] --> R
    E[Global guardrails] --> R
```

<Info>
  Global guardrails can also be managed in [Code Gateway](/en/admin/code-gateway). If you turn off `Inherit global guardrails` at the organizational unit level, that OU is not affected by global guardrails — see [Organization management](/en/admin/organizations#per-ou-guardrails).
</Info>