Implementation Guide for Healthcare Security Engineers

    Small-Cell Suppression for Snowflake MCP Server Queries on Patient Cohorts

    Enforce small-cell suppression at runtime in the MCP execution path. Evaluate each cohort query result against an organization-defined minimum cell size before it reaches the model. Return an allow, redact or block outcome, or escalate for human approval, and record the policy version, threshold and outcome for every decision. HIPAA sets a 'very small' risk standard rather than a numeric minimum, so the threshold is a value your organization must set and justify.

    Evidence to capture for each suppression decision

    • Policy name and version in force at decision time
    • Minimum cell size applied, and the reference to its risk analysis
    • Tool call identity, calling agent and timestamp
    • Query metadata and result metadata, such as dimensions and the number of cells below threshold
    • Outcome: allow, redact, block or escalate
    • Approver identity and decision, where the result was escalated

    Data-layer controls compared with controls in the MCP execution path

    Where it runs

    Data layer: evaluates a query against tables, columns and rows. MCP execution path: evaluates the tool call, the query it produces and the returned result.

    Agent context

    Data layer: sees the query as submitted. MCP execution path: sees the tool name, its parameters and the calling agent.

    Result handling

    Data layer: constrains what the query can return. MCP execution path: inspects what is about to be passed to the model.

    Possible outcomes

    Data layer: masked values, filtered rows, constrained aggregation. MCP execution path: allow, redact, block or escalate for human approval.

    Evidence produced

    Data layer: a record of the query. MCP execution path: a record of the decision made on the result the model received.

    How small-cell risk arises in agent-driven cohort queries

    A count is a small cell when it covers very few people. Even with no direct identifiers present, someone familiar with the population could deduce who those people are. This is the reasoning behind long-standing federal practice on small cell sizes.

    Agents make the problem easier to trigger. A human analyst usually asks a handful of questions. An agent can refine a cohort filter repeatedly, adding a diagnosis, an age band, a facility or a month until the count drops. The more detail each result carries, the closer a set of aggregate tables comes to a file of individual records.

    HIPAA does not give a number to work with. The Privacy Rule offers two de-identification methods. Expert Determination requires a qualified person to find that the risk of identification by an anticipated recipient, alone or in combination with other reasonably available information, is very small. Safe Harbor requires removal of 18 specified data elements. The Privacy Rule does not require a particular approach to reach 'very small'. CMS has stated that it does not publish cell sizes of 10 or less. That is agency practice, not a HIPAA threshold, so treat it as a reference point for your own policy value.

    Choosing the enforcement point

    The path is short. The agent calls an MCP tool with cohort parameters. The MCP server runs a query against Snowflake. The result returns to the server, and the server passes it to the agent and the model. Policy can be enforced at three points: the tool-call parameters, the query that is generated, and the result.

    The result stage is where the minimum cell size must be applied, because only there are the actual counts known. The earlier stages add value as well. Parameter checks can reject filter combinations that are unlikely to produce an acceptable result, and a review of the query shape can catch grouping dimensions that were never intended for agent use.

    Trussed enforces policy in the execution path for tool calls and MCP servers, with allow, redact, block, rate limit and escalate for human approval outcomes. The scope is the individual interaction and tool call. It is enforced at runtime, not reviewed after the fact, and it can be deployed self-managed in the customer cloud so that result content stays inside your environment.

    Implementing a minimum cell size at runtime

    The steps below describe one workable sequence. Order matters, particularly generalizing before suppressing.

    • 1. Set the threshold as a documented policy value: Choose the minimum cell size through an expert or documented risk analysis. Record the methodology and results behind it. Do not borrow a number from CMS practice without that justification.
    • 2. Generalize before suppressing: Combine small categories or reduce table dimensions where you can. Treating at-risk records before tabulation is preferable to suppressing cells, because suppression costs utility.
    • 3. Define the outcome per result: Allow results at or above the threshold. Redact cells below it, which is local suppression. Block results where redaction would leave nothing useful or expose the suppressed value by arithmetic.
    • 4. Check for complementary disclosure: A suppressed cell can be recovered from totals and neighboring cells. In no table should all cases of a line or column fall in a single cell, so evaluate whole result sets, not individual cells.
    • 5. Route edge cases to human approval: Escalate when a legitimate business need requires a result below the threshold. Name the approver role and record who decided.
    • 6. Constrain iterative refinement: Rate limits and limits on filter dimensions slow an agent narrowing a cohort step by step. Per-call enforcement does not by itself detect differencing across queries, so decide how you will address it.

    Tradeoffs and ongoing review

    Every suppression rule trades data utility for privacy. A higher threshold protects more but returns less, and agents may compensate by issuing more queries. Generalizing categories first reduces how often redaction is needed.

    De-identification leaves some residual risk, so review the threshold periodically and keep the supporting analysis current. A per-result check also has limits. It does not remove the need for data-layer controls, and it does not by itself resolve differencing across repeated queries. Treat those as separate design decisions.

    Trussed runs in production with healthcare and pharma organizations. Whether you build this enforcement yourself or adopt a product, the useful test is the same: can the control see the tool call and the result before the model does, and can it prove what it decided?

    Where the threshold is applied

    Label: Agent tool call

    Description: The agent requests a cohort count with filters.

    Label: Pre-execution check

    Description: Parameters and query shape are evaluated.

    Label: Snowflake query

    Description: Data-layer controls apply as configured.

    Label: Result check

    Description: Cell counts are compared with the minimum cell size.

    Label: Outcome and evidence

    Description: Allow, redact, block or escalate, with a decision record.

    See how runtime enforcement and evidence work on MCP tool calls

    Review how policy is enforced in the execution path and how each allow, redact or block decision becomes audit-ready evidence.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.