Section 2 · Prompt Engineering · 20% of exam
2.7 Trust Layer Security and Privacy Features
Exam objective
Identify how the Einstein Trust Layer supports security, privacy, and safer use of Salesforce data with generative AI.
The Einstein Trust Layer is Salesforce's set of controls for handling generative AI requests in a Salesforce context. It helps protect data as prompts move from Salesforce applications to an LLM and responses return. The exam expects students to recognize its security and privacy role, especially when a prompt uses customer or business information. It should be understood as a collection of protections in the prompt journey, not as a replacement for Salesforce permissions or sound data governance.
Protect prompt data
A prompt can include CRM values, user instructions, and retrieved context. The Trust Layer helps apply controls such as data masking for sensitive information, secure data handling, and a zero-retention arrangement with supported model providers so prompt data is not retained or used to train their models. The exact available controls can depend on the feature, model, and release, so administrators must consult current Salesforce documentation and configuration settings.
Data masking is intended to reduce exposure of sensitive values when data is sent for processing. The model may receive placeholders where masked values were present, while Salesforce can use the original values within the trusted workflow as applicable. This does not mean every sensitive field is automatically protected in every design. Admins must know which fields and resources are included, verify what appears in the prompt preview, and avoid passing unnecessary personal or confidential information.
The Trust Layer can also apply safety and grounding controls that help reduce unsafe or unsupported responses. Grounding provides relevant data to improve accuracy; response checks and policies can help constrain content. These controls lower risk but do not eliminate the need to test outputs or establish human review where appropriate.
Preserve platform authorization
Salesforce permissions remain important. Users should only access records and fields they are authorized to see. The administrator must also ensure that prompt templates, actions, and automations execute in a suitable security context. Trust controls do not grant permission to read a record, and they do not automatically decide which user should be allowed to run a template. Review user permissions, sharing, field-level security, the data resource, and the output destination as one design.
Apply governance throughout the lifecycle
Trust is not a one-time toggle. Identify the data used by the prompt, choose an approved model and feature, test with representative and sensitive scenarios, monitor actual usage, and revise when behavior or requirements change. Explain to stakeholders what the controls do and what they do not guarantee. An organization remains responsible for deciding whether a use case is appropriate, whether users need review, and how generated content is retained or acted upon.
When a question asks about protecting prompt data sent to an LLM, consider the Trust Layer. When it asks whether a user can access a Salesforce field or record, check platform permissions as well.
Zero data retention refers to the supported provider arrangement described by Salesforce; it is not a claim that Salesforce itself never processes data or that every model and feature has identical settings. Extra concept: Masking, grounding, and authorization solve different problems: masking reduces exposure, grounding supplies relevant context, and authorization determines what a user or process can access. Official reference: Salesforce Help, Einstein Trust Layer - https://help.salesforce.com/s/articleView?id=ai.generative_ai_trust_arch.htm&language=en_US&type=5
Key takeaways
- The Trust Layer adds privacy, security, and safety controls to the AI prompt journey.
- It complements Salesforce permissions and data governance; it does not replace them.
- Verify the actual data, model, configuration, and output path for a use case.
