Implementation Guide | Healthcare

    Governing Model Context Protocol Toolsets in Google ADK Agents That Access PHI

    To govern MCP toolsets in a Google ADK agent that touches PHI, use four layers: toolset configuration, tool filters, callbacks, and the MCP server boundary. Configuration and filters decide which tools exist for the agent. They do not decide what a given call sends or returns. Per-call decisions (allow, redact, block, rate limit, escalate for human approval) need runtime enforcement, and 164.312(b) audit evidence should record each decision.

    Four control layers and what each can enforce

    1. 1

      Toolset configuration

      Decides which MCP servers the agent connects to and how it authenticates. It can keep an unapproved server out of the agent. It cannot see what any call contains.

    2. 2

      Tool filters

      Decides which of a server's tools the model can see. It narrows the choice set. It cannot distinguish a permitted call with harmless arguments from the same tool called with PHI it should not carry.

    3. 3

      Callbacks

      Run around individual tool calls inside the agent application, so they can inspect arguments and results. They depend on the application code and apply only to agents that include them.

    4. 4

      MCP server boundary

      Covers server version, credential scope and network egress. Nothing inside the agent can correct a vulnerable or over-privileged server.

    Audit evidence to specify for 164.312(b)

    • Inventory of registered MCP servers and tools per agent, with versions and the credential each server uses
    • The tool-level policy in force at the time of each call
    • Per-call record of the tool, arguments, results and the decision applied
    • Redactions applied, and what category of PHI was removed
    • Human approval requests with approver, time and outcome
    • Blocked and rate-limited calls, retained as evidence of enforcement

    Where PHI crosses the tool boundary

    An ADK agent can be given MCP toolsets that expose EHR queries, patient records, scheduling and other PHI-bearing systems. The model discovers those tools, chooses among them and writes the arguments for each call. It then receives the results in its context. Each of those steps is decided at runtime, not at deployment.

    This gives two crossing points that a security review should mark explicitly. The first is outbound: PHI in the conversation or in earlier tool results can end up in the arguments of a tool that was never scoped for it. The second is inbound: a tool returns more PHI than the task needs, and that data now sits in model context where it can be repeated, summarized or passed to a different tool.

    A third concern is trust. Tool results are data from another system, and the model reads them as part of its working context. A January 2026 federal request for information on AI agent security lists indirect prompt injection, where a model interacts with adversarial data, among the recognized risks. Treat every tool result as untrusted input, including results from internal systems.

    Why the server boundary needs its own controls

    NVD records CVE-2026-16481, an SSRF and credential-exfiltration flaw in the cloud-healthcare-fhir-fetch-page tool of Google's MCP Toolbox for Databases. Depending on configuration, it leaks either the end-user's token or the broader service-account access token, which could potentially expose PHI and secondary Google Cloud services. The affected range is versions from 0.19.1 up to but excluding 1.5.0. Whether any particular ADK deployment uses this tool is a question for its own inventory.

    The lesson is structural. Agent-side filters and callbacks cannot repair a server that leaks its own credentials. Three controls belong at the server boundary: patching tracked against advisories, credentials scoped to what each tool needs, and network egress limits. The difference between an end-user token and a broad service-account token sets how far exposure extends if something goes wrong. Tools that accept URLs or other free-form targets deserve specific testing with unexpected arguments.

    Define a tool-level access policy

    Write the policy as a table with one row per tool, not one row per server. A server-level decision is too coarse when a single server exposes both a scheduling lookup and a full record fetch. Each row should answer the same questions, and the answers become the rules that every layer implements.

    • Identity: Which agent, server and tool this row covers, and the credential the server uses to reach the underlying system.
    • PHI in arguments: Which identifiers or clinical content may be sent to this tool, if any.
    • PHI in results: Which fields may return to the model, and which must be redacted before they do.
    • Decision: Allow, redact, block, rate limit or escalate for human approval, with the condition that triggers it.
    • Approval owner: Who approves escalated calls, such as record writes or bulk reads, and what they see when they decide.
    • Server hygiene: Approved server version, credential scope and permitted network destinations.

    Placing decisions: build time versus runtime

    Build-time controls are cheap, predictable and reviewable. Register only approved servers, filter each toolset down to the tools the use case needs, and give every server the narrowest credential that works. These steps shrink the surface and should come first.

    They do not make a decision about a specific call. Redacting PHI from an argument, blocking a request that targets an unexpected URL, rate limiting a bulk read and pausing a record update for human approval all depend on the content of that call. That content exists only at runtime. Callbacks can cover this inside the agent, with the tradeoff that policy logic is then spread across application code and each agent team has to implement it consistently.

    Enforcing in the execution path, outside the agent code, centralizes the policy. Trussed works this way: it enforces policy on individual interactions and tool calls, including MCP servers, with allow, redact, block, rate limit and escalate for human approval, enforced at runtime, not reviewed after the fact. Trussed runs in production with healthcare and pharma organizations. Whichever mechanism you choose, the tradeoff to evaluate is the same: consistency of enforcement against added latency and operational ownership.

    Where PHI crosses the MCP tool boundary

    Model to tool

    Tool selection and arguments are generated at runtime and can carry PHI.

    Tool to model

    Results from EHR, record and scheduling tools return PHI into model context.

    Tool to MCP server

    The server's credentials and network reach define the exposure if it is compromised.

    Every call

    Policy decision and outcome recorded as audit evidence.

    See what enforcement and evidence look like per tool call

    Trussed enforces policy in the execution path on individual interactions and tool calls, and produces audit-ready evidence mapped to HIPAA, including 164.312(b).

    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.