You are here:
LLM Data Masking (Available for Non-Agentforce Generative AI) Control
Automatically identifies and masks sensitive data (PII, PCI, and so on) within a prompt before it is transmitted to an external Large Language Model (LLM).
Control Name
Einstein Trust Layer - LLM Data Masking (Available for non Agentforce gen AI)
Control Overview
Automatically identifies and masks sensitive data (PII, PCI, and so on) within a prompt before it is transmitted to an external Large Language Model (LLM).
Description
Uses pattern matching and machine learning to replace sensitive entities (for example, names, emails, credit card numbers) with placeholders, which are de-masked only after the response returns to Salesforce.
Recommended Configuration
Enable "LLM Data Masking" in Einstein Setup. Customize specific masking policies for all relevant entities (SSN, phone, email) based on your privacy and compliance requirements.
Security Impact
Prevents sensitive data from being shared with LLM providers, to support compliance with global privacy regulations like GDPR and CCPA.
Business Impact
Enables the safe adoption of gen AI for customer-facing and internal workflows without risking data residency violations or intellectual property leaks.
Security Risk If Not Configured
Unmasked sensitive data (PII/PHI) is sent to the LLM provider (for example, OpenAI, Anthropic) in plaintext, potentially violating data processing agreements and organizational policies.
Threat Scenarios
Data Leakage: A user inadvertently incorporates sensitive fields in the prompt.
Estimated CVSS Score Range
Critical (9.0–10.0).
Risk Impact Considerations
Risk is extreme for orgs in regulated sectors (Finance, Health) where sending a single unmasked record to an external cloud can trigger a mandatory breach notification.
Higher Risk When
Gen AI users incorporate data from sensitive fields (for example, PII/ PCI data) in their prompts in most of the use cases.
Low Risk When
Zero data retention (ZDR) is technically enforced with the external LLM providers, or internally hosted model is used to train the LLM developed by your company.
Business and Integration Considerations
Masking can occasionally reduce the contextual accuracy of the LLM if too much data is obscured. Careful testing of prompt templates is required to balance security and utility.
Security Health Review Guidance
Security Health Review audits the Einstein Trust Layer Setup to confirm that data masking is enabled.
Who Is Impacted
Data privacy officers, AI developers, admins, and any users using Prompt Builder or Einstein Copilot features.
