Comparison: healthcare AI security

    Microsoft Purview Sensitivity Labels vs Runtime Policy for Stopping PHI in SharePoint MCP Server Responses

    Sensitivity labels classify and protect content, and they are most reliable inside applications that read and honor them. Runtime policy inspects each MCP tool-call response before it reaches the model and applies allow, redact or block, and it records the decision. Labels do not replace that enforcement point, and runtime policy does not replace classification or permissions. A healthcare team exposing SharePoint to agents needs both, plus tight access control on the identity the MCP server uses.

    Follow one request from agent to model and back

    1. 1

      1. Agent issues a tool call

      The request carries the user, agent and task context. Any policy that depends on who is asking needs this context at this point.

    2. 2

      2. MCP server authenticates to SharePoint

      The server acts as the end user or as a service identity. This choice determines which permissions, and which label rights, are evaluated.

    3. 3

      3. SharePoint evaluates access and returns content

      Site and item permissions apply here. A label is metadata on the item and may or may not travel with what is returned.

    4. 4

      4. MCP server shapes the response

      Custom server code decides what is extracted and returned, such as text, fields or file contents, and in what format.

    5. 5

      5. Response travels back to the agent

      This is the control point for inspecting content before it enters the model context.

    6. 6

      6. Model receives the content

      Once PHI is in context it can be summarized, repeated or passed to later tool calls.

    What sensitivity labels do on this path

    A sensitivity label marks content with a classification and can attach protection such as access restrictions. Its value is that classification is applied consistently to files and sites at rest, and that applications which understand labels can enforce the associated rules. For PHI in SharePoint, labels answer the question of what the content is and how it should be handled.

    The limit is where enforcement happens. A label restricts access only when something in the path reads it and acts on it. An MCP server is custom or third-party code, and an agent client may not be label-aware. Whether a labeled item is withheld, returned as plain text, or returned with its label stripped depends on the identity the server uses, how the server extracts content, and what the consuming client honors. If the server runs under a service identity with broader access than the end user, the permissions evaluated are the service identity's, not the user's.

    Treat a label as a classification input, not as proof that a response was filtered. Test what happens to labeled content at retrieval with your specific server and client, and do not assume the answer from the label configuration alone.

    It also helps to keep terms apart. The security labels used in health IT exchange standards for tagging documents are a different concept from Purview sensitivity labels.

    What runtime policy must do on a tool-call response

    Runtime policy operates on the response itself, at the moment of the tool call. To stop PHI it has to do four things: receive the response before the model does, detect PHI in the returned content, choose an action, and record the decision.

    The actions differ in cost and risk. Allow passes the response unchanged and is appropriate when no PHI is found or the use is permitted. Redact removes or masks detected identifiers and lets the rest through, which keeps the task working but depends on detection being complete. Block withholds the whole response and is the safest option for high-risk categories, at the price of failed tasks. Each action should be set per data category and per requester, not globally.

    Two tradeoffs deserve direct measurement. Detection accuracy has two failure modes: a false negative lets PHI through, and a false positive blocks or mangles legitimate content. Added latency accrues on every tool call, and inspection depth affects it. Neither can be assumed. Test on representative PHI from your own content and measure latency in your own environment.

    Redaction is also not the same as HIPAA de-identification. Masking detected identifiers does not make output de-identified under HHS de-identification guidance, so do not describe redacted output that way unless a defined method has been applied.

    Labels and runtime policy compared by control point

    The two approaches act on different things. The table summarizes where each applies and where each stops.

    • What it acts on: Labels: the content item, at rest and in label-aware applications. Runtime policy: each individual tool-call response in transit.
    • Decision basis: Labels: classification assigned to the item. Runtime policy: what the response actually contains, plus requester, agent and task context.
    • Typical gap: Labels: enforcement depends on a label-aware reader, and a service identity may see more than the user. Runtime policy: detection can miss PHI or flag clean content, and it does not fix over-broad source permissions.
    • Available actions: Labels: protection and access restriction as configured. Runtime policy: allow, redact or block per response.
    • Evidence produced: Labels: proof of classification and configured protection. Runtime policy: a record of each response decision and the action taken.
    • Relationship: Complementary. Labels and permissions reduce what can be retrieved, and runtime policy governs what is returned to the model.

    HIPAA audit evidence from each approach

    The HIPAA Security Rule establishes national standards to protect electronic protected health information created, received, used or maintained by a covered entity or business associate. The technical safeguards sit in 45 CFR 164.312, which includes the audit controls provision at 164.312(b). HHS ties safeguard selection to the risk analysis, so exposure of ePHI to AI agents through MCP servers belongs in that documentation. These details are current as of November 2026.

    The two approaches produce different evidence. Labels show that PHI-bearing content was classified and that protection was configured. That supports a claim about how data is governed at rest. It does not show what a given tool call returned to a model. Runtime policy can produce a per-call record: who or what requested the content, which source item it came from, what the policy decided, and which action was applied. That record speaks to what PHI actually moved into model context, which is the question an auditor will ask about an agent.

    Before relying on either, read the full text of 164.312(b) and map each evidence item to it. Also confirm whether a site should hold PHI at all. Some organizations prohibit it; one HHS agency's privacy assessment states that PHI is not authorized in its SharePoint Online sites. That is an agency-specific rule, not a HIPAA requirement, but it shows that site-level policy is worth checking before any control is designed.

    When evaluating a design, ask five questions. Does the MCP server act as the user or as a service identity? Does any control inspect each response before the model receives it? What is recorded per tool call, and can it be exported for audit review? What are the measured false-negative and false-positive rates and latency on your own content? Does a redaction claim rest on a defined de-identification method or only on masking?

    Enforcement in the execution path: one example

    Trussed is an example of the runtime layer described above. It enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. Actions include allow, redact, block, rate limit and escalation for human approval. Policy is enforced at runtime, not reviewed after the fact.

    That enforcement generates continuous, audit-ready evidence mapped to HIPAA, among other frameworks, which supports the 164.312(b) mapping work described earlier. Trussed can be self-managed in the customer cloud, with configurable retention and data residency, and it runs in production with healthcare and pharma organizations.

    It does not replace labels, permissions or DLP. Those controls reduce what can be retrieved, and enforcement in the path governs what is returned to the model and records the decision.

    One tool call, four control points

    Permissions

    Decide what the calling identity can read in SharePoint.

    Sensitivity labels

    Classify and protect content at rest and in label-aware applications.

    DLP

    Applies content rules in the services and locations it covers.

    Runtime policy

    Evaluates each tool-call response and allows, redacts or blocks it before the model sees it.

    Review your SharePoint MCP response path

    Walk through identity, labels, DLP and response-level enforcement for one tool call with our team, and see what evidence each control produces.

    Talk to an Expert