Custom detections and library

When the built-in Detection Models do not cover what you need caught, such as an internal project codename, a customer ID format or a topic you want to keep out of AI tools, add a custom detection model. You can install a pre-built one from the Detection library or build your own with the Detection Model builder.

Go toQuilrAI ConsoleGovernDetection ModelsData RisksCustom

Detection types

Type	How it matches	Good for
Precision	Deterministic patterns (regex)	Identifiers with a fixed shape: employee IDs, ticket numbers, internal hostnames
Semantic	Meaning of the text rather than its exact form	Content that is phrased in many ways
Intent	What the user is trying to do, described in plain language and steered by examples	Requests such as competitor comparisons or asking for legal advice

The Custom view shows how many models of each type you have, with a card per model offering Edit & test and Delete.

Detection library

The Detection library drawer holds about 140 pre-built Precision and Intent models.

1. Open Detection library from the Custom view.
2. Search or filter for what you need.
3. Select Try to test a model against sample text.
4. Select Install to add it to your tenant. Installed models appear in the Custom view (View in Custom).

Detection Model builder

The builder turns a plain-language brief into a detection model.

1. Open the Detection Model builder from the Custom view.
2. Write a brief describing what to detect, for example "internal project codenames that start with PRJ- followed by four digits".
3. Choose Quick or Advanced mode, then select Start building.
4. The builder suggests a model type (Precision for fixed patterns, Intent for behaviour) and drafts it.
5. Test the draft against sample text, including near misses that should not match.
6. Save the model when the results are right.

Builds run in the background and are listed under builder runs, so you can leave the page and come back.

Test before you enforce

Add negative examples that are close to the positive ones. Then use the new detection in a policy on Monitor first, review the findings for a few days, and only then move to redact or block.

Use a custom detection

Custom models appear in the Custom group of the data type picker on the Data & Adversarial Risks card in the Policy Engine, so a policy can give them their own action, threshold, stage and scope. LLM Gateway apps can also hold app-level custom detections; see LLM Gateway custom detections.

From red team findings

Red Teaming reports can suggest custom detections that would stop the attacks they found, with patterns and a suggested action. Add them here, then use them in a policy. See Turn findings into detections.