Pattern-based and field-based LLM data masking
Key Takeaway
LLM data masking hides sensitive data from the LLM in two ways. Pattern-based masking detects it in the prompt text, and field-based masking masks the values of merge fields that have Shield Platform Encryption or a data classification tag. Both types are disabled for agents.
Large language model (LLM) data masking is an Einstein Trust Layer feature. It finds sensitive data in a prompt, such as personally identifiable information (PII), and replaces it with placeholder text before the prompt goes to the LLM. There are two masking types, pattern-based masking and field-based masking, and they differ in how they find the sensitive data.
How masking replaces sensitive data
Masking replaces each piece of sensitive data with placeholder text based on what the data represents, for example a name or an email address. The LLM sees only the placeholders. After the LLM returns a response, the Einstein Trust Layer demasks the response, so the user sees the actual data.