Guardrails

Define the rules and boundaries within which the agent must operate. Guardrails describe what the agent must not do, should avoid, or must stay within while achieving its goal. This ensures the agent behaves safely, ethically, and appropriately.

Content Filters

Content Filters help control the types of content the agent can receive and generate. Select a predefined filter profile based on the agent's intended use and risk level.

Select a Content Filter profile (Analysis, Balanced, or Strict) to define the level of protection applied to agent inputs and outputs across content categories such as Hate / Insults / Misconduct, Sexual / Violence, and Prompt Attack, with filtering levels configured as Low, Medium, High, or N/A.

Analysis: Minimal filtering designed for internal analytics, investigation, or support use cases.

Balanced: Moderate filtering suitable for most enterprise scenarios.

Strict: Maximum filtering recommended for customer-facing, regulated, or high-risk environments.

Filtering	Analysis		Balanced		Strict
Category	Input	Output	Input	Output	Input	Output
Hate / Insults / Misconduct	Low	High	Medium	High	High	High
Sexual / Violence	Medium	High	Medium	High	High	High
Prompt Attack	High	N/A	High	N/A	High	N/A

Sensitive Data

Sensitive Data guardrails help protect confidential information by allowing agents to identify and anonymize sensitive data during interactions. In addition to built-in data types, you can define custom regex patterns to detect organization-specific data formats

Built-in Data Types: Includes predefined sensitive data categories that are automatically anonymized. These protections are always enabled and cannot be modified or removed. The list shows included types to view the supported data types.

Custom Regex Pattern: Define additional regular expression (regex) patterns to detect organization-specific identifiers, account numbers, customer references, or other proprietary data formats that are not covered by the built-in sensitive data types.

To create a custom regex pattern:

Click on the + Add Regex Pattern button.
In the Name field, enter a meaningful name that identifies the purpose of the pattern.
In the Regex Pattern field, enter the regular expression that should be used to detect the required data format.
From the Action list, select how the agent should handle matching content:
- Anonymize: Replaces the detected value with an anonymized version.
- Block: Prevents the detected content from being processed or returned.
From the Scope list, select where the pattern should be applied:
- Input: Applies the rule only to content submitted to the agent.
- Output: Applies the rule only to content generated by the agent.
- Input and Output: Applies the rule to both incoming and outgoing content.
Click Save Pattern to add the regex pattern to the Sensitive Data configuration.
Click Cancel to close the window without saving the pattern.

Once saved, the custom regex pattern is added to the Sensitive Data list and becomes active according to the configured action and scope.

Word Filters

Word Filters help prevent unwanted language from being used in agent interactions. You can enable a profanity filter or define custom words and phrases that should be blocked.

Profanity Filter

Profanity Filter automatically detects and blocks common profanity. Select where the filter should be applied:

Input & Output: Applies profanity filtering to both content submitted to the agent and content generated by the agent.
Output Only: Applies profanity filtering only to content generated by the agent.

Blocked Words

Use Blocked Words to define specific words or short phrases that should not be allowed in agent interactions.

Enter one or more words or phrases (up to three words each).
Add multiple entries by separating them with commas.
Select where the blocked words should be applied:
- Input & Output Blocked: Blocks specified words in both incoming and outgoing content.
- Input Only Blocked: Blocks specified words only in content submitted to the agent.
- Output Only Blocked: Blocks specified words only in content generated by the agent.
After configuring the required settings, click Create Agent to save the configuration and continue with agent creation, or select Cancel to discard your changes.

Denied Topics

Denied Topics allow you to control whether the agent can respond to specific categories of sensitive or regulated subjects. For each topic, you can choose to Allow or Deny agent interactions.

When a topic is set to Deny, the agent will not provide guidance, recommendations, or responses related to that subject. When a topic is set to Allow, the agent can process and respond to requests related to that area.

The following predefined topics are available:

Medical Advice: Covers requests related to symptoms, diagnoses, treatments, medications, medical tests, and healthcare recommendations.
Legal Advice: Covers requests involving laws, regulations, contracts, legal rights, disputes, liabilities, and legal interpretations.
Investment Advice: Covers recommendations related to investments, asset allocation, financial planning, and investment strategies.
Banking and Credit Advice: Covers guidance related to loans, mortgages, credit products, refinancing, lending, and borrowing decisions.
Tax Advice: Covers tax-related guidance, filing decisions, deductions, tax liabilities, and tax planning strategies.
Mental Health Crisis: Covers requests involving self-harm, suicide, emotional crises, or situations requiring immediate professional assistance.

Select the appropriate Allow or Deny option for each topic based on your organization's compliance, safety, and business requirements.

Blocked Message

Blocked Messages define the text displayed to users when a request or response is blocked by the configured guardrails. You can customize separate messages for blocked inputs and blocked outputs to provide clear feedback to users.

Input: Specifies the message displayed when content submitted to the agent is blocked due to a guardrail policy violation.
By default: Input content is blocked due to policy violations.
Output: Specifies the message displayed when content generated by the agent is blocked due to a guardrail policy violation.
By default: Output content is blocked due to policy violations.

Best Practices

Apply the most restrictive Content Filter profile that meets your business requirements.
Block only terms that are inappropriate or prohibited for your use case. Excessive use of blocked words may prevent the agent from responding effectively to legitimate requests.
Review guardrail settings regularly. As business requirements, compliance policies, and agent use cases evolve, verify that content filters, denied topics, sensitive data rules, and blocked words continue to provide the appropriate level of protection.
Test guardrails before deploying agents to production. Validate content filters, sensitive data rules, denied topics, and blocked messages using representative test scenarios to confirm that the configured protections behave as expected.

FAQ

What is the difference between Input and Output filtering in Content Filters?

Input filtering evaluates content submitted to the agent and helps prevent unsafe or inappropriate requests from being processed. Output filtering evaluates content generated by the agent and helps prevent unsafe, sensitive, or non-compliant responses from being returned to users.

Which Content Filter profile should I choose?

Select Analysis for internal analytics or support scenarios, Balanced for general enterprise use cases, and Strict for customer-facing or regulated environments that require stronger content controls.

What is the difference between Built-in Data Types and Custom Regex Patterns?

Built-in Data Types provide predefined sensitive data detection that is always anonymized and cannot be modified, while Custom Regex Patterns allow you to create additional rules to detect organization-specific data formats.

Why should I deny certain topics?

Denying high-risk or regulated topics helps reduce the risk of the agent providing information in areas that may require professional expertise, regulatory oversight, or human intervention.