Implementation Guide | Healthcare

    Preventing the Microsoft 365 MCP Server from Emailing PHI to Domains Without a BAA

    Treat the MCP tool call as the enforcement point. Build a recipient-domain allowlist from your BAA inventory, evaluate every send-mail or draft tool call against it before execution, and allow, redact, block, or escalate for human approval based on recipients and content. Keep native Microsoft 365 controls as additional layers, and record each decision as evidence to support HIPAA 164.312(b) audit controls.

    Design the decision at the tool call

    1. 1

      1. Identify the tool and caller

      Match the tool name, agent identity and requesting user to the policy that applies.

    2. 2

      2. Resolve recipient domains

      Check every to, cc and bcc address against the BAA-backed allowlist.

    3. 3

      3. Inspect content

      Run PHI detection on the subject, body and attachment content that the call carries.

    4. 4

      4. Choose the action

      Allow when all recipients are allowlisted. Redact when content can be safely masked. Block when PHI is headed to a non-allowlisted domain.

    5. 5

      5. Escalate edge cases

      Route ambiguous or high-risk calls to a named approver, with the content and recipients shown.

    6. 6

      6. Rate limit

      Cap send volume per agent so a looping or compromised agent cannot send at scale.

    Record evidence for 164.312(b) audit controls

    • Timestamp, agent identity and requesting user
    • Tool invoked and the policy version applied
    • Recipients and resolved domains, with allowlist match result
    • Content and attachment inspection result, including any redaction made
    • Decision taken: allow, redact, block or escalate
    • Approver identity and outcome for escalated calls

    Map the PHI-to-email exposure first

    When an AI agent reaches Microsoft 365 through an MCP server, the mail capability becomes a tool the model can call. Any tool that sends a message, or creates a draft that a later step sends, is a path for PHI to leave the organization. The agent may pull PHI from a calendar entry, a document, a prior thread or a user prompt and place it in the body or an attachment.

    Start with an inventory. List every MCP tool exposed to each agent that can send mail or create drafts. Record the Microsoft Graph permissions behind the app registration that the MCP server uses, and whether those permissions apply to one mailbox or to many. Then record, for each tool, which parameters are available to inspect before execution: recipients, subject, body and attachments. A policy layer can only evaluate what the tool call carries.

    Apply least privilege at this stage. If an agent only needs to read mail, it should not hold send permissions. If it needs to draft, consider whether a human sends. Narrowing the app registration reduces what any later policy must catch, but it does not decide which recipients are acceptable. That decision needs its own control.

    What each layer can and cannot see

    Each control sits at a different point in the path, and that position determines what it can evaluate. No single layer covers every question, so the design should state plainly what each one contributes.

    • Graph permission scoping and least-privilege app registrations: Limits which mailboxes and actions the MCP server can use. It does not evaluate individual recipients or content per call.
    • Exchange mail flow rules: Act on messages in transit and can match on sender, recipient domain and message properties. They see the composed message, not which tool or agent produced it or the policy that applied.
    • Purview DLP: Detects sensitive content in messages and documents against configured rules. It operates on content after composition and does not carry the agent's tool-call context.
    • Tool-call policy at runtime: Evaluates the tool invoked, its recipients and its content before the action executes, and can allow, redact, block, rate limit or escalate. It depends on the quality of the allowlist and PHI detection behind it.

    Tie the recipient allowlist to BAA status

    HHS states that the HIPAA Rules permit a covered entity to disclose PHI to a business associate if it obtains satisfactory assurances, in the form of a contract or other written arrangement, that the business associate will appropriately safeguard the information. That written arrangement is the business associate agreement. The allowlist should therefore be derived from your BAA inventory, not maintained as a separate list that drifts away from it.

    Give the allowlist an owner and a change process. Each entry should reference the agreement it relies on, the date it was reviewed, and who approved the entry. Domain-level entries are simple to enforce but coarse. Decide whether a BAA covers an entire domain or specific addresses, and encode that distinction if it matters.

    Whether a given recipient is a business associate is a legal determination, and a BAA is not necessarily the only basis on which a disclosure may be permitted. Involve privacy counsel so that the policy reflects the organization's position, and so that exceptions are defined deliberately rather than discovered in production.

    Rollout, tradeoffs and where Trussed fits

    Begin in a monitoring posture for a short period, reviewing which calls would have been blocked, then move to enforcement. Expect false positives from PHI detection and tune the thresholds with the teams that send mail. Human approval adds latency and requires staffed approvers, so reserve it for calls where redaction or blocking would be wrong. Review the allowlist on a fixed schedule, because a lapsed or terminated BAA should remove the domain from the list promptly.

    Keep native controls running. Mail flow rules and DLP remain a useful backstop for messages that originate outside the agent path, and permission scoping limits the blast radius if the runtime layer is misconfigured.

    Trussed is the runtime layer for the tool-call enforcement described here. It enforces policy in the execution path across agents, tools and MCP servers, with actions that include allow, redact, block, rate limit and escalate for human approval, scoped to individual interactions and tool calls. Enforcement produces audit-ready evidence mapped to HIPAA, and it can be deployed self-managed in the customer cloud with configurable retention. Trussed runs in production with healthcare organizations. The design in this guide stands on its own regardless of which tool implements the runtime layer.

    Where policy sits in the MCP email path

    Agent request

    The assistant decides to send or draft an email.

    MCP tool call

    Tool name, recipients, body and attachments are visible here.

    Runtime policy

    Domain allowlist and PHI checks produce allow, redact, block, or escalate.

    Microsoft 365 mail flow

    Native DLP and mail flow rules act on the resulting message.

    Evidence record

    Each decision is stored for later examination.

    See policy enforced on individual MCP tool calls

    Walk through how recipient allowlists, PHI checks and approval routing are enforced at runtime, and what evidence each decision produces.

    Request a Demo