Salesforce Says Masking Is Off. A Salesforce Architect Says It's On. Someone's Wrong.
Salesforce's own Help documentation states, twice on the same page, that data masking through the Einstein Trust Layer is disabled for agents. A Salesforce architect, writing under his own name with his title disclosed, has published that every A2A call is intercepted by the Trust Layer specifically to mask PII before it reaches the LLM.
Both of these are public. Both are specific. They cannot both be describing the same behavior of the same product.
This has real deployment consequences. Whichever version an architect believes changes whether they think their org needs additional PII controls in front of an agent, or already has them covered by the platform.
What each source actually says
Salesforce Help, current page. The Trust and Agentforce article states that data masking through the Einstein Trust Layer is disabled to improve the performance and accuracy of agents, and separately, under its own heading, that pattern-based and field-based data masking for LLMs is disabled for agents. This is the current, live version of the page: its table of contents references Multi-Agent Orchestration (Beta), so it reflects the current product, not a stale carryover from an earlier release.
Salesforce Ben, published February 13, 2026. In an architecture guide on designing A2A systems, the article states that every A2A call is intercepted by the Trust Layer, framed as PII masking meant to keep sensitive customer data from reaching the LLM. The article's modified timestamp matches its published timestamp to the second, meaning it hasn't been revised in the five months since it went up. The author, Varun Misra, is credited in his bio as a Director and Technical Architect at Salesforce. That gives the claim real weight inside the ecosystem.
Salesforce Dictionary, updated May 16, 2026. A third source states flatly that you cannot fully disable masking as a feature. This is the detail that turns two conflicting sources into a genuine muddle rather than a simple question of who's right. Now there are three claims about the same mechanism, each one public and dated, and no two of them agree.
Why the distinction matters
An architect who starts with the Salesforce Ben piece could reasonably conclude the Trust Layer is already masking sensitive fields before anything reaches an LLM or a downstream agent, and treat that as a control they can rely on. Start instead with Salesforce Help, and the picture flips: masking is off, and the responsibility for an equivalent control sits with the architect, not the platform. Those are two different architectures. Only one of them matches what Salesforce's own current documentation says is actually happening, and the org that builds on the wrong one is missing a control it thought it had.
What to do about it
- Treat Salesforce Help as authoritative over ecosystem content when the two disagree, and verify the actual setting in your own org rather than assuming either source describes your specific configuration.
- If masking is off, as Salesforce's documentation states, build PII handling into the agent's own instructions and grounding sources. Don't rely on an implicit platform guarantee that the platform's own docs say doesn't exist.
- Ask your AE or a Salesforce architect directly which of these three claims is accurate for A2A specifically, and get the answer in writing. Even a Salesforce-titled author has published the opposite of what Salesforce's own help page says. That alone is reason enough not to take either claim on faith.
Close
When official documentation and a named practitioner inside the same company's orbit disagree about a basic yes-or-no fact, the responsible move is to verify in the sandbox rather than pick whichever version is more convenient, and to stop treating vendor-adjacent content as a single, coherent source of truth. If someone from Salesforce can point to the update that reconciles this, this post will be updated to reflect it.
Sources credited: Salesforce Help (Trust and Agentforce, and Data Masking Limitations in Agentforce), Varun Misra and Salesforce Ben (How to Design Salesforce Agent-to-Agent (A2A) Architecture, published February 13, 2026), Salesforce Dictionary (Salesforce Einstein Trust Layer: The Complete 2026 Guide, updated May 16, 2026).