PII masking

PII masking protects sensitive data before it reaches a model. The guardrail detects selected entity types and replaces their values with opaque tokens.

To configure it, open Manage Agent → Guardrails. Turn on Sensitive Data Masking, then select the gear in that row.

Turning on Sensitive Data Masking.

When this guardrail is off, content reaches the model unmasked. Enable it whenever the agent handles PII.

Choose a masking provider

A provider detects PII in agent content. Choose one based on its coverage, accuracy, and latency.

Provider	What it does	Use it when
AWS Comprehend	Managed ML detection with broad entity coverage	You need the widest catalog and can use a managed ML service
Azure Text Analytics	Managed ML detection with EU identifiers	You need EU-specific identifiers
Google Cloud DLP	Managed inspection with custom infoTypes	You need a large catalog or custom infoTypes
Regex	Pattern matching without ML	You need low latency for structured values, such as SSNs and card numbers
Off	Does not mask content	No PII flows through the agent

Choosing a masking provider.

The selected provider controls the available entity types. Regex exposes six types. Google Cloud DLP exposes 133 types. Managed-provider coverage reflects your tenant configuration.

Set the entity rule

Use Entity rule to choose which detected entities are masked.

• Deny list masks only selected entity types. Add each type that needs protection.

• Allow list passes selected entity types through unchanged. It masks every other detected type.

Use Select all or Clear to update every entity. Use Clear group to reset one group. The page shows how many entity types are masked.

Setting the entity rule and selecting entity types.

Entity groups

Entity groups vary by provider. Use Search to find a specific type.

Regex (6 entities)

• Identity: US SSN (SSN) and Passport (PASSPORT_NUMBER)

• Contact: Email (EMAIL) and Phone (PHONE)

• Financial: Bank account (BANK_ACCOUNT_NUMBER) and Card number (CREDIT_DEBIT_NUMBER)

Google Cloud DLP (133 entities)

Examples include:

• Technical: IP address (IP_ADDRESS)

• Other: Age (AGE) and Auth Token (AUTH_TOKEN)

• Country-specific identifiers: Argentina DNI, Australia Medicare, and Australia Tax File Number

Select Save to apply provider or entity-selection changes.

Understand tokenization

PII masking uses tokenization. The platform replaces a detected value with an opaque token before a model receives the content. The mapping is stored in a secure vault. A permitted downstream step can restore the original value when needed.

Selected entities are tokenized. Unselected entities bypass tokenization and reach the model unchanged. Tokenize all detected entities unless a clear requirement needs the original value.

Exclude an entity type only when:

• A downstream system requires the original value.

• A secure downstream step de-tokenizes the response.

• Detection for that type obscures required context.

Tokenization requires a compatible executor version. When the agent uses an unsupported version, the entity list is read-only. Update the executor version in Runtime & Versioning.

Review your configuration

Review the entity list when the agent’s responsibilities change. Every excluded type reaches the model in clear text.

For the parent-level overview, see Guardrails.