> ## 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.

# Custom AI services

> Add an AI service the catalogue does not know, such as an internal tool or a new vendor: its host, the paths to read, how its requests are read, how a reply is scanned, and which detectors run.

A custom AI service covers a tool that no public catalogue knows: your company's own
chatbot, or a vendor that is not listed yet. A company's own chatbot is never
recognised on its own, so this is where you add it. You give Blindsight the domain.
Devices then read the traffic to it and apply the same inspection as for any other AI
service.

> [!WARNING]
>
> Adding a service means devices in your workspace terminate TLS to that domain,
> from their next policy check. Check that the domain is the one you mean. Every
> addition is recorded in the [audit trail](https://docs.blindsight.io/compliance/audit-trail.md).

## Who can do what

| Permission | Lets the user |
| - | - |
| `endpoint.ai_services.add` | Add a service, and send a test prompt. Held by `admin` and `endpoint-admin` by default. |
| `endpoint.ai_services.teach` | Choose a ready-made format or teach one from a sample. Held by `admin` only by default, because the format decides which part of a request is scanned. |
| `endpoint.ai_services.manage` | Change who may use a service, its detectors and its reply setting, and delete it. |
| `endpoint.discovery.view` | See custom services and their health. |
| `runtime_security.firewall.configure` | Allow company certificates on devices, in the same step as an add or from the switch beside the added services. |
| `endpoint.ai_services.private_routes` | Turn on reading internal services on private addresses, and, while it is on, add a service or change the names one covers. Held by `admin` only by default, and not available to MSP operators. |

## Add a service

On **Endpoint → AI services**, select **Add service**. The dialog has three
steps.

**1. Which service to add**

Enter the **Domain** people reach the service at, for example
`chat.example.com`. A URL is fine: the domain is kept, and a path in it fills the
paths field.

* **Which names it covers**: **This host and its subdomains**, or **This host
  only**. A bare domain also covers its subdomains, so `example.com` covers
  `chat.example.com`.
* **Only these paths** (optional): one path per line. A star matches any run of
  characters, for example `/v1/chat/completions*`. Leave it empty to read every
  path on the host.
* **Name** (optional): the vendor or product name. It only changes what the
  inventory shows.

**2. How this service is treated**

Choose **Inspect** or **Protect**. Both read the traffic. They differ in what
happens when something is found.

* **Inspect**: personal data is pseudonymized before a prompt leaves the
  machine and prompt injection is scored. Nobody is blocked from using it.
* **Protect**: the same inspection, and a request that trips a rule can be
  stopped. It is treated like an AI service the company runs.

Then choose **Who may use it**: **Everyone**, or **Only these groups**. With
groups, everyone outside them is blocked from the service, and each group gets
the action you choose. Limiting a service to groups needs the manage permission.

**3. How its requests are read**

Choose **Not sure** to read every request the generic way. This is right for
most services. Or choose a ready-made format. A format can be taught later from
a sample.

The new row opens in the **Catalogue** so the next step is in front of you.

A host that cannot be inspected is refused, with the true reason. That covers a name
that is not a domain, an IP address, a registry suffix, a host Blindsight never
intercepts and a host the agent bypasses. A cloud AI endpoint outside the three
below is refused for the same reason: add your own chatbot's domain instead.

### Cloud model endpoints by exact name

You can add your own **Azure OpenAI**, **Azure AI Foundry** or **Amazon Bedrock**
endpoint, by its exact name only:

| Provider | The name you add |
| - | - |
| Azure OpenAI | `<resource>.openai.azure.com` |
| Azure AI Foundry | `<resource>.services.ai.azure.com` |
| Amazon Bedrock | `bedrock-runtime.<region>.amazonaws.com` |

The name is held exact, with no wildcard: the provider's other endpoints, and other
people's resources on the same domain, are never read. The dialog says what will be
read before you save. Devices read every request that any app or script on the
machine sends to that endpoint, not only your chatbot's, and personal data in it is
handled like any other AI service. Bedrock requests are signed, so they are never
changed: a request whose personal data would be replaced is stopped instead. An
endpoint is read only once your workspace has added it. Approving or protecting a
cloud model endpoint that nobody added never makes devices read it.

### Company certificates

An internal service often uses a certificate your company issued. Devices do not
reach such a service until an administrator allows company certificates. Adding a
service changes nothing about trust on its own. The dialog says so, and offers
**Also allow company certificates on devices (recorded in the audit trail)** to a
person who holds `runtime_security.firewall.configure`. The same choice is the
**Allow company certificates** switch beside the added services.

When it is on, devices also accept the certificates your company installs on its
computers when they check a site, so an internal AI service with a company
certificate is reached. It applies to every site devices inspect, so turning it on
asks you to confirm. Each change is recorded in the audit trail, and removing a
service does not turn it off.

### Internal services on private addresses

On Windows, an internal chatbot is often reached on a company-network address. The
switch **Read internal AI services on private addresses**, beside the added
services, lets Windows devices route the company-network address of each service
you added through Blindsight, so the chatbot is read there too. Only the exact hosts
added with **Add service** are routed, never a network a VPN keeps to itself. Turning
it on shows a warning first, needs `endpoint.ai_services.private_routes`, and is
recorded in the audit trail, as is turning it off. A desktop app that does not
support it ignores the switch.

### Hosts that match one name only

Two kinds of host cover their own name only, never their siblings, and the dialog
says **This host only** with the reason:

* A **cloud application host**, such as an API Gateway, a load balancer or a
  `cloudapp.azure.com` name. Its siblings belong to other people.
* A name under a domain **shared by many tenants**. A rule for it covers that name
  only, never the other tenants.

## Reach of a service you added

Adding a host does not mean Blindsight can read what is sent to it. Open
**Endpoint → AI services**, then a service you added in the **Catalogue**. Its row says
what devices did with it:

| Reach | What it means |
| - | - |
| **Reached** | Devices reached it and read the requests. The line shows the counts. |
| **Seen, nothing readable** | Requests reached the service but carried nothing the endpoint could read. |
| **Not reached yet** | No device has reached it. |
| **Too few to show** | In a workspace that shows only totals, a count under the minimum is not shown. |

Choose **Inspect** or **Protect** as described in [Add a service](#add-a-service).
Requests to Amazon Bedrock are signed over their body, so a Bedrock service can be
added, but its prompts are read and blocked, never redacted. Under **Protect**, a
prompt that would be redacted is blocked instead. See
[Cloud model endpoints by exact name](#cloud-model-endpoints-by-exact-name).

## The service detail

Open the service in the **Catalogue** to see its sections.

### Where

The host, whether subdomains are covered, and the paths read. For example
`chat.example.com and its subdomains, 2 paths`. Devices read only those paths. An
empty list reads every path.

### Format

The format says where in a request the prompt sits. It reads **Generic**, the name
of a ready-made format, or **Taught**.

* **Not sure (generic)**: the platform looks for text that reads like a prompt.
* **Template**: pick a ready-made format. The places read are listed under
  **Read from**.
* **Teach from a sample**: paste a request body from the service (JSON or a form
  body), then click the strings the person types. The sample stays in your
  browser. It is not sent, stored or logged. Only the paths you choose are saved.

Teaching needs `endpoint.ai_services.teach`. Without it the section says so.

### Detectors

**Detectors** shows what runs for this service. A detector the workspace has off
stays off. The section holds:

* The per-service **Personal data** and **Prompt injection** settings. See
  [Per-service settings](https://docs.blindsight.io/endpoint/ai-services.md#per-service-settings-customize-for-this-service).
* **Secrets and keys**. To stop protecting a kind of personal data for this
  service, set it to **Keep** under **Data types** in **Customize for this
  service**. It covers the same categories, such as email addresses, phone
  numbers, person names, postal addresses, card numbers, bank account numbers
  and national and tax identifiers.
* **When something is found**: **Redact** or **Block**. **Block** is offered only
  under **Protect**.
* **Scan a streamed reply once it ends**. This is the reply verdict. The reply
  reaches the person untouched. A verdict is recorded afterwards, and nothing is
  rewritten.

On-device personal data protection follows **Inspect** or **Protect**.

### Who

**Everyone**, or **Only these groups**. With groups, "Only people in the listed
group may use this service. Everyone else is blocked."

### About

When it was added and by whom, an owner, a description and a classification:
**Public**, **Internal**, **Confidential** or **Restricted**.

### Health

Four figures. A figure the platform cannot answer reads as a dash with the reason,
never as 0.

| Figure | Meaning |
| - | - |
| **Last prompt** | When a prompt was last read, and when a reply was last reported. |
| **Unread requests, 7 days** | Requests seen but not read. |
| **Unclaimed endpoint shapes** | Requests to this host that carry no prompt the service reads. If one carries prompts, teach the format from a sample. |
| **Devices applying paths and replies** | Devices on agent 1.7.0 or later. |

### Prove it

Select **Send a test prompt**. The page shows a sentence with a code in it, and
asks you to type the whole sentence into the service from any enrolled device.

The page then checks every 5 seconds for up to 15 minutes and shows:

* **Seen on a device**: which device, and the result, such as **Scanned**,
  **Blocked**, **Protection paused on that device**, **Seen, not inspected** or
  **Seen without a host**.
* **Read from**: the field the prompt was read from. An agent that does not report
  it says **Not reported by this agent**.
* **Verdict**: **Allowed**, **Redacted** or **Blocked**.
* **Reply**: **Reply scanned**, **Reply observed, not scanned** or **No reply seen**.

If nothing arrives in time, the page lists why, most likely reason first. Up to
three tests can be open at a time.

## Older agents

Path scope, the reply verdict and the per-service settings are applied by agent
**1.7.0 and later**. An older agent reads every path on the host and follows the
workspace reply setting, the workspace pseudonymization and the workspace
injection setting. Detectors in the scan, who may use the service and the test
prompt work for every agent. The page says how many devices run an older agent and
never reports an older fleet as applying something it ignores.

## See also

* [AI services](https://docs.blindsight.io/endpoint/ai-services.md)
* [Endpoint permissions](https://docs.blindsight.io/endpoint/permissions.md)
* [Release notes](https://docs.blindsight.io/release-notes.md)


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