Skip to main content
Control and security · Guardrails

Sensitive data is protected
before it reaches the model

Guardrails check what goes into and comes out of the model. Input and output are checked with rule-based detection, AI evaluation and prompt injection detection, and blocked or pseudonymized according to policy. Every decision is recorded in the guardrail log.

Contact sales
Settings screen of the “Customer Data Shield” guardrail
Settings screen of the “Customer Data Shield” guardrailReal product screen
Guardrails

Sensitive data is protected
before it reaches the model

Rule-based detection, AI evaluation and prompt injection detection inspect every input and output. Depending on policy, data is blocked or pseudonymized, and every verdict is recorded in the guardrail logs.

Chat inputTry editing it
Action
What the model receives

Where checks run

Detection runs on every request at each enforcement point. Hover over or select an item in the diagram to see its decision.

Sample

13 enforcement points

Chat inputModel outputAttachmentsText in imagesTool resultsMemoryDB extractionFilesScheduled tasksAPICode gatewayAsk the manual
13personal data types6actionsConfidential document similarity detection
Real product screen
Contact details and card numbers reach the model as pseudonyms
What changes

Before and after adoption

Before

Employees may paste resident registration numbers or account numbers as they are.

After

Guardrails detect 13 types of personal data and either block it or replace it with pseudonyms before it is sent to the model. When the user views the answer, the original values are restored.

Before

Instructions hidden in documents or tool results can change how the AI behaves.

After

Prompt injection detection checks input, attachments and tool results, and also finds AI instructions hidden in synced documents.

Before

It is hard to find out later what was blocked and where.

After

Decisions at all 13 enforcement points are recorded in the guardrail log.

How it works

Four steps

01Choose how to check

Turn on what you need from personal data rules, AI evaluation, discriminative model evaluation and prompt injection detection.

A guardrail's check method settings
Key features

Guardrail features in detail

13 personal data types

Also detects personal data in Korean formats, such as resident registration numbers, account numbers, card numbers and driver's license numbers.

6 handling methods

Choose from pseudonymize, mask, hash, redact, block and “log only.”

AI evaluation and discriminative model evaluation

A model reads the context of each sentence and checks it again.

Prompt injection detection

Finds hidden instructions. A separate model can be assigned to make decisions on attachments and tool results.

Pause after repeated violations

When it is on, a user is warned after 5 violations and blocked for 30 minutes after 10.

Confidential document similarity

Finds content similar to confidential documents you have registered.

Document sensitivity classification

Classifies documents into the standard 4 levels (Public · Internal · Confidential · Restricted) or the 3 levels of Korea's N2SF (C Classified · S Sensitive · O Open).

Image screening

Checks images and the text in them.

Tool call checks

Checks every tool call to confirm that the user asked for that action.

Admin settings

Settings each company configures

  • Create guardrail policies and set their enforcement points.
  • Set the handling method for each item. At input, you can also warn the user and have them enter a reason.
  • Turn on Pause after repeated violations.
  • Choose the document sensitivity classification scheme: the standard 4 levels or the 3 levels of Korea's N2SF.
13Personal data types detected
13Guardrail enforcement points
6Handling methods

Figures reflect the product as of October 2026.

Related features

Related pages

Permissions and share reviewAnswers within each user's permissions and shares only after approvalCode gatewayApplies company policy to developers' coding toolsAdministration and auditManage users, sign-in, audit logs and retention periodsSecurity and complianceKorean financial and public-sector guidelines
FAQ

Frequently asked questions about guardrails

Doesn't pseudonymization make answers less accurate?

The model builds its answer from the relations between the pseudonyms, and the pseudonyms are changed back to the original values when the answer is shown to the user. Only pseudonyms are sent to the model.

Can I log detections without blocking the content?

Yes. Choose “Log only,” and the original text is sent unchanged while only the detection results are recorded in the guardrail log.

Where do guardrails apply?

At 13 enforcement points, including chat input, model answers, attachments, text in images, tool results, memory, database extraction, files, scheduled tasks, the API, the code gateway and Ask the guide. You set whether they apply at each point.

Which items in Korea's financial and public-sector guidelines do they address?

See the Security and compliance page for the check items each control addresses.

See it with
your own documents

In a consultation, we connect a few of your documents so you can check the answer quality for yourself. Send us your questions and our team will get back to you.

Contact salesRequest security documentation
Product videosReal product screens · Example company ACME Logistics · 16 videos