> ## 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.

# Topic rules

> Flag or block prompts about a subject you describe in plain language, whatever words people use.

A topic rule flags or blocks prompts about a subject you describe in a sentence.
You do not list keywords. Blindsight matches prompts by what they are about, so a
prompt that uses different words for the same subject is still caught.

Use topic rules for what a pattern or a data type cannot express, for example
"asking for administrator, root or sudo access without approval" or "drafting
messages to candidates about salary offers".

## Add a topic rule

Topic rules live in a policy. Open **Endpoint → Policies**, open a policy, and go
to the **Topics** section.

**1. Describe the topic in a sentence**

Write what the prompts are about, in plain language. Be specific, so ordinary
prompts are not caught. The editor tells you when a description is too short or
too vague.

**2. Choose the action**

* **Flag** keeps the prompt and records it.
* **Block** stops the prompt for everyone the policy covers.

A new topic always starts as **Flag**.

**3. Add example prompts (optional)**

Short example prompts help Blindsight recognize the topic. Each example is a full
sentence, for example "Can you give me sudo on the build server?".

**4. Save the policy**

The save bar lists the topics added, removed and changed. Devices the policy
reaches pick up the change like any other policy change.

> [!TIP]
>
> Start with **Flag** and check the flagged events on **Endpoint → Events** for a
> few days. Switch the topic to **Block** once the matches are the ones you expect.

After you save, Blindsight prepares recognition examples for the topic, so it is
recognized in other wordings and languages. The topic works straight away. While
the examples are being prepared, a line under the topic says so. If they cannot be
prepared yet, the topic still works and Blindsight tries again. You can also ask it
to try again from that line.

## Who a topic rule applies to

A topic rule belongs to its policy. It reaches the people, groups and devices the
policy is assigned to, the same way as the policy's other rules. See
[Where to edit policy](https://docs.blindsight.io/endpoint/policies.md#where-to-edit-policy).

Changing the words of a topic makes it a new rule with a new ID. Changing only its
action or its examples keeps the same rule.

## Matching by meaning

Matching is by meaning, not by exact words. A topic matches prompts about the same
subject however they are worded, so you do not need to list every phrasing.

## Start from the policy description

When a notice or a link takes you to a policy's **Topics** section, the editor can
add a first topic filled in from the policy's description. Blindsight keeps the
subject of the sentence and drops the wording around it. For example, "Employees
must not use AI assistants to gain administrator access" becomes a topic about
gaining administrator access. Check the text, choose the action and save.

## What people and admins see

* **The person** whose prompt is blocked sees that it was stopped, with the topic
  as the reason, for example **Prompt about asking for administrator access**.
* **Admins** see the event in **Endpoint → Events** with the reason `topic_rule`
  and the topic that matched.
* **Your SIEM** receives the event with the reason `topic_rule` and the ID of the
  topic rule, if you use [syslog export](https://docs.blindsight.io/endpoint/syslog-export.md). The wording of
  the topic is not sent.

## When topic rules do not run

Topic rules need prompt inspection. If prompt inspection is off for the workspace,
the **Topics** section says that topic rules will not run yet. Your topics are kept,
and they start working when inspection is switched on. See
[Protections](https://docs.blindsight.io/endpoint/protections.md).


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