Implementation Guide | Healthcare

    Break-Glass Access for AI Agents on PHI MCP Servers: Time-Boxed Elevation with Evidence

    Implement break-glass access for AI agents as policy enforced on each tool call in the execution path, not as a standing credential. Each elevation needs a written trigger, a named human approver, a narrow tool and data scope, a fixed expiry that is enforced at runtime, and an evidence record that ties the approval to every elevated call and to the data returned or redacted.

    Elevation lifecycle

    1. 1

      1. Request

      The agent or its operator submits a trigger reference, justification, requested tools, data scope and duration.

    2. 2

      2. Escalate for approval

      Policy holds the request until a named human approves, narrows or denies it.

    3. 3

      3. Activate

      The approved scope and expiry are recorded as policy state bound to the agent identity.

    4. 4

      4. Enforce per call

      Every tool call is evaluated against the active elevation before it reaches the MCP server.

    5. 5

      5. Expire

      At expiry, or on revocation, further elevated calls are denied at the enforcement point.

    6. 6

      6. Review

      The record is reviewed after use and checked against the trigger and scope.

    Evaluating a runtime enforcement approach

    • Does it decide per tool call, or only at the session or credential level?
    • Does an elevation expire at the enforcement point, and what does the agent receive on the next call?
    • Does each record tie approver, justification, scope and expiry to specific calls and to data returned or redacted?
    • Can the evidence stay in your own cloud, with retention you control?
    • Can the agent bypass the enforcement point with a direct connection to the MCP server?
    • Can the human approval step be scoped to one request rather than granted for a whole session?

    What HIPAA asks of emergency access, and why agents change the picture

    As of November 2026, the access control standard in 45 CFR 164.312 requires technical policies and procedures for systems that maintain ePHI so that access is allowed only to those persons or software programs that have been granted access rights. The wording covers software programs, so it is not limited to human users. The same standard lists an emergency access procedure as a required specification: establish, and implement as needed, procedures for obtaining necessary ePHI during an emergency. Unique user identification is also listed as required, and audit controls are the subject of 164.312(b).

    Subpart C defines access as the ability or means necessary to read, write, modify, communicate or otherwise use any system resource. An agent tool call that retrieves or changes ePHI on an MCP server plausibly fits that definition. This is a reading of the text, not agency guidance on agents, so confirm it with counsel.

    The regulation does not prescribe an approver, a maximum duration or a log format for emergency access. Those are design decisions that belong in your risk analysis. That is where traditional break-glass breaks down for agents. Shared or standing emergency credentials make actions hard to attribute to a single identity, and after-the-fact review cannot stop an over-broad call that has already returned PHI.

    Define the pattern before choosing tooling

    Start with triggers. Write down the clinical, operational and incident-response conditions under which an agent may request elevation, and tie them to your organization's emergency access procedure. A trigger should be specific enough that a reviewer can later decide whether it was met, for example a named incident ticket or a documented clinical escalation.

    Next, define the elevation itself. Bind it to a distinct agent identity, a named tool or small set of tools on the PHI MCP server, a data scope such as a patient context or record type, and an expiry. Set a default duration and a maximum in policy. Avoid reusing a human break-glass account for the agent, because it undermines attribution.

    Require a justification and a named human approver. Neither is mandated by the regulation, so treat them as design choices that make the emergency access procedure defensible. Finally, assume the agent can be manipulated. Federal AI agent security discussion notes that agent systems can take autonomous actions that affect real systems and may be hijacked. Keep elevation narrow, and evaluate every elevated call on its own instead of trusting the session as a whole.

    Express elevation as policy over individual tool calls

    A credential answers who may connect. Break-glass needs an answer to a narrower question: should this call, with these arguments, against this resource, be permitted now? That is a runtime decision, and it should be made in the execution path between the agent and the PHI MCP server.

    Five outcomes cover most needs. Allow calls that fall inside the approved tool and data scope. Redact fields the emergency does not require, such as identifiers outside the patient context. Block calls to tools or resources outside the scope. Rate limit elevated calls so that a compromised or looping agent cannot pull bulk records within the window. Escalate for human approval when a call goes beyond the approved scope or touches a higher-sensitivity tool.

    Expiry should be enforced at the same point. If the elevation lapses, the next call is denied by policy, and the agent receives a clear denial it can surface to the operator. Relying on credential revocation alone leaves a gap between the expiry time and the moment the credential is actually invalidated. Also decide how the enforcement point is positioned relative to the server. If the agent can connect to the MCP server directly, per-call policy can be bypassed, so network and routing controls must make the enforced path the only path.

    The evidence record each elevation must produce

    The audit controls standard is 164.312(b). The regulation does not specify the fields of a break-glass record, so define them from what a reviewer will need to reconstruct: who asked, why, who approved, what was permitted, what happened and when it ended. Confirm the operative audit controls wording in the current regulation when you finalize the schema.

    The record should be written by the enforcement point at decision time, not assembled later from application logs. That keeps the approval, the scope and each call in one linked chain under a single elevation identifier. Set retention in policy, based on your risk analysis and counsel's advice, and store the records where they cannot be altered by the agent or the service it calls.

    • Requesting agent: The distinct agent identity and the operator or system that initiated the request.
    • Trigger and justification: The condition invoked, the ticket or incident reference and the stated reason.
    • Approver: The named human, the decision, any narrowing of scope and the approval timestamp.
    • Scope and duration: Permitted tools, data scope, default and maximum duration, and the actual expiry time.
    • Tool calls made: Each call with its timestamp, arguments, policy decision and the elevation identifier.
    • Data outcome and expiry event: What was returned, what was redacted, and the event that ended the elevation, whether lapse, revocation or denial.

    Agent break-glass lifecycle

    Trigger

    A written condition and justification start the request.

    Approval

    A named human approves a defined scope and duration.

    Per-call enforcement

    Each tool call is allowed, redacted, blocked, rate limited or escalated.

    Expiry and evidence

    Elevation lapses at the enforcement point and leaves a durable record.

    Test your break-glass pattern against runtime enforcement

    See how per-call policy, human approval and audit-ready evidence apply to agents on PHI MCP servers.

    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.