Implementation Guide | Healthcare

    Keeping protected health information (PHI) out of agent actions on the SharePoint Model Context Protocol server

    Treat the SharePoint MCP server as a tool surface with four PHI exposure points: request parameters, retrieved content, returned metadata, and write or share actions. Place a policy decision at each point, in the execution path, so every tool call is allowed, redacted, blocked, or escalated before PHI reaches model context, a downstream tool, or an output. Record each decision as evidence that supports HIPAA audit controls under 164.312(b).

    Applying allow, redact, block, and escalate at runtime

    1. 1

      Allow

      The call matches an approved operation and scope, and no PHI is detected where none is expected. Record the decision and pass the call through.

    2. 2

      Redact

      The task is legitimate but the response contains PHI the task does not need. Remove or mask the fields before delivery and record what was removed.

    3. 3

      Block

      The request targets out-of-scope content, or a write or share action would move PHI to a disallowed destination. Stop the call and return a defined refusal.

    4. 4

      Escalate

      The call is high-impact or ambiguous, such as sharing a document that contains PHI. Hold it for human approval and record the approver and outcome.

    Where PHI enters and leaves agent actions

    An MCP server exposes a set of tools to an agent. For SharePoint, these typically cover operations such as search, file read, list enumeration, file write, and sharing. The exact set depends on the server you deploy, so the first implementation step is an inventory of every operation the agent can call and the arguments each one accepts.

    PHI can move in both directions on each call. On the way in, the agent's own arguments can contain PHI, for example a patient name inside a search query. On the way out, the server returns content and metadata that become part of model context the moment they are delivered. A file name such as a patient surname and procedure is PHI-bearing even when the file body is never opened. Write and share operations then carry content back out, to another library, another recipient, or a downstream tool that receives the agent's output.

    This four-point framing is a design model for analysis, not a regulatory taxonomy. Its value is that each point can have its own policy and its own evidence.

    Why identity permissions leave a per-call gap

    SharePoint permissions answer a question about identity: may this account access this item? That control is necessary. It does not decide whether a specific request or response is appropriate for the task the agent is performing. An agent running under a service identity with broad library access can legitimately read a document and still pull far more PHI into context than the task needs.

    The design consequence is that identity-level access control and call-level policy do different jobs. Permissions bound what is reachable. Call-level policy evaluates what is being asked and what is being returned, on each invocation. Existing label and data loss prevention configurations should be assessed against MCP-mediated access directly, by testing representative tool calls, rather than assumed to apply.

    Detection and redaction tradeoffs

    Redaction is only as good as detection. Structured content, such as list columns with defined fields, lends itself to deterministic rules because the location and format of identifiers are known. Free-text documents are harder: identifiers appear in varied forms and context determines whether a string is PHI. Contextual detection handles this better but costs more compute per call.

    Because enforcement sits in the call path, detection adds latency to every tool call. Test structured fields and free-text documents separately, on your own SharePoint content, and measure both detection quality and added delay. Decide in advance how the policy behaves when detection is uncertain. For write and share actions, failing closed is the more defensible default. For low-risk reads, a lower-confidence redaction may be acceptable. Do not adopt vendor accuracy or latency figures without reproducing them on your own data.

    Evidence that supports HIPAA audit controls

    The HIPAA Security Rule sets national standards for ePHI that a covered entity or business associate creates, receives, uses, or maintains. Technical safeguards sit in 45 CFR 164.312, and 164.312(b) is the audit controls provision. The Privacy Rule is in Subpart E of the same part. A January 2025 proposal to update the Security Rule was published, and as of November 2026 it should be treated as a proposal, not current law.

    What follows is a design recommendation, not a statement of what the regulation prescribes. Have counsel and your compliance team confirm the mapping against the regulation text. For each tool call, capture enough to show what the agent was allowed to do and what it did:

    The identity and agent that made the call, the tool and operation requested, the policy that applied, and the decision taken. Add a record of what was returned, redacted, or written, and for escalations, who approved and when. Store the record durably, make retention configurable, and make it exportable so an auditor can review it without access to the live system.

    Keep the scope honest. Per-call evidence shows individual interactions. It does not by itself reconstruct a multi-step workflow.

    Evaluating a runtime enforcement approach

    Five questions separate in-path enforcement from after-the-fact review. Does the decision occur before content reaches model context? Which stages can it evaluate: request, retrieved content, metadata, write and share? Which decisions are supported, and how are escalations handled and recorded? What evidence is produced per call, and can it be exported? What are the measured detection accuracy and added latency on your content?

    Trussed is one example of this approach. It enforces policy in the execution path across models, agents, tools, and MCP servers, with allow, redact, block, rate limit, and escalate-for-human-approval decisions scoped to individual interactions and tool calls. Enforcement generates audit-ready evidence mapped to HIPAA, and it can be deployed self-managed in the customer cloud with configurable retention and data residency. It runs in production with healthcare organizations. Test it against the same five questions you apply to any alternative.

    PHI exposure points on a SharePoint MCP tool call

    Request parameters

    Queries, paths, and filters the agent sends may carry PHI or target PHI-heavy libraries.

    Retrieved content

    File bodies and search snippets enter model context once returned.

    Returned metadata

    File names, titles, authors, and library paths can identify patients.

    Write or share actions

    Content written, copied, or shared can move PHI outside the task scope.

    Test enforcement on your own tool calls

    See how in-path policy decisions and audit-ready evidence work on agent tool calls against PHI-bearing content.

    See How Evidence Is Produced