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

# Protections

> Choose which of the three endpoint protections devices run: prompt injection detection, personal data protection and Shadow AI. Each works on its own.

A workspace decides what its devices do with AI traffic by switching
three protections on or off. They are independent: any combination
works, so you can run Shadow AI alone, or protect personal data without
looking for injection.

Open **Endpoint → Configurations → Protections**. It is the first item
in the rail, and the only place the three are switched.

## The three protections

| Protection | When it is on | When it is off |
| - | - | - |
| **Prompt injection detection** | A redacted copy of each prompt is checked for injection and jailbreak attempts. | Those attempts go undetected. |
| **Personal data protection** | Personal data is redacted or pseudonymized before prompts and uploads leave the device. | Prompts and uploads reach AI services unchanged. |
| **Shadow AI** | Devices record which AI services people use, including ones nobody approved. | Unapproved AI services are not seen. |

A new workspace starts with all three on. Existing workspaces keep all
three on until you change them.

At least one protection stays on. The last switch that is on cannot be
turned off; to stop all of them, remove the agent. See
[Uninstalling the agent](https://docs.blindsight.io/endpoint/uninstall.md).

### Choices under each switch

Sub-choices appear only under a switch that is on.

* **Personal data protection**
  * **Redact** removes the value. **Pseudonymize** swaps it for a
    stand-in the person can get back, with the mapping kept on the
    device.
  * **Find names on the device** runs the names model locally. Labels,
    retention and scope are under **Personal data details**.
* **Shadow AI**
  * **Unapproved AI** is **Record only**, **Protect** or **Block**.
    This is the posture towards services nobody has decided on. See
    [Shadow AI triage](https://docs.blindsight.io/endpoint/shadow-ai/triage.md). **Protect** needs
    injection detection or personal data protection on, because it
    reads prompts.

A sentence under the switches describes the combination you have chosen,
so you read the outcome and not only the switches.

## What each combination does

| Combination | What a device does |
| - | - |
| All three | Reads prompts to approved services, protects personal data, flags injection, and records every AI service in use. |
| Injection and personal data, no Shadow AI | Reads prompts, protects and flags, and does not record which services people use. |
| Shadow AI only | Records which AI services people use. **Prompts are never read.** |
| Personal data only | Protects personal data in prompts and uploads. Nothing is checked for injection. |
| Injection only | Checks a redacted copy of each prompt for injection. Personal data is not changed. |

### Shadow AI only means no prompt is read

With injection and personal data both off, a device decrypts nothing. It
sees which AI services people connect to and nothing they send. The one
exception is a site you set to **Block**: the connection is ended only to
show the block page, and the request is never read or forwarded.

### Turning personal data off covers everything

With **Personal data protection** off, secrets and your custom rules stop
redacting too, not just personal data. The prompt reaches the AI service
unchanged. The copy sent to Blindsight for injection scoring is still
redacted on the device, so turning this off never sends raw personal data
to Blindsight.

## Confirm before turning one off

Saving a change that turns a protection off asks you to confirm, and
lists in plain words what will stop. The change is recorded in the audit
trail, one record per changed field. Turning a protection off is recorded
as high severity.

## Devices that do not honour a switch yet

A device on an older agent version follows what the platform tells it for
Shadow AI and for injection detection. An older agent keeps redacting
personal data whatever you choose, which is the safe direction. The page
shows how many devices are in that state, and they follow your choice
after they update.

## Licence

A protection your licence does not include shows **Not in your plan**
instead of a switch, and the platform refuses a request to turn it on.

## Who can change it

Changing protections needs `endpoint.protections.configure`. Admin and
`endpoint-admin` hold it by default. See
[Endpoint permissions](https://docs.blindsight.io/endpoint/permissions.md).

## Setting it during Getting Started

The first question of **Connect people's devices** in Getting Started is
**What should devices do?**, with the same three switches. All three
start ticked. It saves as you go and shows the same page as above, so the
two cannot disagree. The rest of the steps follow your choice: with
Shadow AI off, the first-AI-service milestone is replaced by the first
protected prompt.

## The API

```bash theme={null}
curl $API_BASE/api/runtime-security/dlp/protections \
  -H "Authorization: Bearer $TOKEN"
```

`GET` returns the three protections, their sub-choices, which are
licensed, and how many devices do not yet honour a switch. `PUT` takes
any of them and saves in one transaction. It refuses turning all three
off, refuses **Protect** when both prompt-reading protections are off,
and refuses turning on a protection your licence lacks.


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