Compliance Guide: Healthcare

    Google ADK and Vertex AI Agent Engine under the HIPAA BAA: what is covered and what you still control

    Whether an ADK agent on Vertex AI Agent Engine falls within Google Cloud's HIPAA BAA depends on your executed BAA and Google's current covered-products list. The agent framework and the managed runtime do not settle it. Check every service in the data path against that list before PHI reaches the agent. Even where a service is covered, the BAA addresses Google's role as a business associate. Prompt and tool-call handling, access scoping and audit evidence under 164.312(b) remain your responsibility.

    Runtime controls your organization must supply

    • PHI handling in prompts and tool calls: decide which data may enter a prompt or a tool argument, and redact or block the rest before it leaves your boundary.
    • Tool-call restrictions: allow only approved tools for each agent and use case, with scoped credentials and least-privilege access to PHI.
    • Human approval: escalate high-consequence actions, such as writing to a record or sending PHI externally, for review before execution.
    • Data-path scoping: keep PHI out of any model, tool or store you have not confirmed as covered or contracted.
    • Evidence retention: retain prompts, tool calls, outputs, access events and enforcement decisions on a defined schedule.
    • Change review: repeat the scope analysis, with compliance, security and legal sign-off, when a model, tool or connector changes.

    What the BAA settles and what it does not

    HHS states that a covered entity or business associate may use a cloud service to store or process ePHI, provided it enters into a HIPAA-compliant BAA with the cloud service provider and otherwise complies with the HIPAA Rules. Using a cloud service provider to maintain ePHI without a BAA is a violation. The BAA is therefore a precondition for PHI reaching an agent.

    The same HHS position makes clear that the BAA is not the whole answer. The customer must still comply with the HIPAA Rules, and any service level agreement should be consistent with the BAA. A cloud provider acting as a business associate is directly liable for its own uses and disclosures of PHI, but that liability does not transfer the customer's obligations to the provider.

    For ADK and Agent Engine, the contractual question is narrow: which specific Google Cloud services are covered by your executed BAA, as shown on Google's current covered-products list? Record the date and version of the list you consulted. Treat coverage as a per-service determination, not an inference from the fact that an agent is built with a Google framework or runs on a managed Google service.

    Mapping an agent deployment against BAA scope

    An agent is not one service. It is a set of components, and each one needs its own scope decision. The table lists the question to answer for each component and who owns the answer.

    • Agent runtime: Is the managed runtime on the covered-products list today? Google's documentation and your BAA answer this. Record the date checked.
    • Model calls: Is each model endpoint the agent calls a covered service? Agents can switch models, so check every one in use, not only the default.
    • Tool and API calls: Does the tool create, receive, maintain or transmit ePHI for you? If so, it is a separate business associate question with its own BAA.
    • Memory and session data: Where is state stored, is that storage covered, and how long is PHI retained? Customer policy decides retention and access.
    • Logging: Do logs capture prompts, tool arguments or outputs containing PHI? If so, the log store is a PHI repository and needs the same scope check and access controls.
    • Customer-side controls: Redaction, access scoping, approvals and evidence retention are yours to supply regardless of BAA coverage.

    Tools, external APIs and the subcontractor chain

    HHS defines a business associate as a person or entity that creates, receives, maintains or transmits PHI on behalf of, or for the benefit of, a covered entity, directly or through another business associate. HHS also notes that no BAA is required with an app developer that does not create, receive, maintain or transmit ePHI on your behalf. That test applies to every tool an agent can call.

    This is where agent deployments tend to widen the compliance boundary. A tool call can send PHI in its arguments to a service that sits outside your Google Cloud BAA. If that provider receives or maintains ePHI for you, you need a BAA with it, and BAAs also run between business associates and their subcontractors. If the provider has no such agreement, the safe design keeps PHI out of that path.

    For each tool, MCP server or external API, document whether PHI can reach it, who operates it, and whether a BAA is in place. Re-run this review whenever a tool, model or connector is added, because an agent's reachable surface changes with its configuration.

    Audit controls under 164.312(b) for autonomous agents

    Section 164.312 of 45 CFR is titled Technical safeguards, and audit controls are the standard at 164.312(b). The section was last amended in 2013. The HHS proposed rule to strengthen the cybersecurity of ePHI, published in January 2025, was still listed as a proposed rule as of November 2026, so it is not current law. It is worth tracking, since the proposal includes documenting and annually reviewing compensating controls.

    Take the exact requirement wording from the current eCFR text, then decide what your agents must produce to meet it. For an agent, the activity that matters is not only user logins. It includes the prompt as submitted, what the agent decided to do, each tool call with its arguments and response, the output returned, and the identity and permissions under which each step ran.

    An auditor will reasonably ask whether those records exist for PHI-touching interactions, whether they are retained for a defined period, who can access them, and whether they show enforcement decisions rather than only activity. Define this evidence set in advance. Records assembled after an incident are weaker than records generated as each interaction happens.

    Enforcing policy in the execution path

    Controls that depend on developers remembering to apply them tend to be inconsistent across agents. The alternative is to enforce policy where interactions happen, so that decisions are made at runtime rather than reviewed after the fact. That means a layer in the path of model calls and tool calls that can allow, redact or block based on whether PHI is present and whether the destination is permitted.

    Trussed is one way to do this. It enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and human-approval escalation applied to individual interactions and tool calls. The enforcement generates audit-ready evidence mapped to HIPAA, and it can be deployed self-managed in your own cloud with configurable retention and data residency. Trussed runs in production with healthcare organizations.

    An enforcement layer is your own control, not a substitute for the BAA analysis. Validate its decisions and evidence outputs against your HIPAA policies, and confirm what BAA and evidence terms apply to any component that handles PHI. The approach is useful only if the underlying scope decisions are already documented.

    Where PHI can enter or leave an ADK agent

    Prompts and outputs

    User input, retrieved context and model responses.

    Model calls

    Each model endpoint the agent invokes.

    Tool and API calls

    Arguments sent and responses returned, including third-party and MCP tools.

    Memory and session data

    Any state persisted between turns.

    Logs and evidence

    Records that support audit controls.

    Define the runtime controls before PHI reaches an agent

    Talk through your agent data paths, tool-call restrictions and the audit evidence your HIPAA program needs.

    Talk to an Expert