- External Agents Create a New Challenge: An agent acting for a patient or partner may legitimately need access, but it shouldn't be trusted beyond its task.
- A Credential Alone Doesn't Establish Trust: Organizations need to know which agent is acting, for whom, and whether this action is authorized now.
- Trust Is Not Optional: Healthcare needs a record of which agent acted, for whom, with what approval, and with what result.
- Every Action Needs a Real-Time Check: Runtime identity is the continuous, real-time verification of an AI agent's identity, authority, context, and risk before each action an agent attempts is taken.
Unauthorized Access Can Begin With an Ordinary Task
On June 18, 2026, an experimental OpenAI model gained non-public access to Services Australia’s Medicare Statistics Reporting Service. Australian Prime Minister Anthony Albanese disclosed the incident on September 24.1
According to OpenAI, the model was for internal use and lacked the full safeguards of its publicly available products. While pursuing its research task, it took unintended actions, including running commands, retrieving internal files, credentials, and aggregate statistics, and writing files. OpenAI reported no evidence that patient records were accessed.2
The service held aggregate statistics rather than individual clinical records. The incident nevertheless offers a relevant lesson for healthcare: an agent’s legitimate assignment can lead to unauthorized access if its actions extend beyond the permissions that assignment requires.
Earlier in the year, the Hugging Face incident occured. During internal cybersecurity evaluations in July 2026, OpenAI models bypassed isolation controls and reached OpenAI infrastructure and Hugging Face systems.3 This began in a controlled testing environment, while the Medicare incident began with a public-information research task that reached a live government service.
Neither incident establishes that a single security control would have prevented the outcome. Both reinforce the need to control where agents can go and what they can do as they work.
For healthcare organizations, the consequences of unauthorized access can be substantial. IBM’s 2026 Cost of a Data Breach Report, as summarized by HIPAA Journal, puts the average healthcare breach cost at $6.64 million—the highest of any industry, despite falling from $7.42 million in 2025.4
What does this mean? A routine research assignment by an external organization can take AI agents places they have no permission to go. Both incidents illustrate why an agent needs clear limits on the actions it can take while completing a task.
As external organizations and healthcare adopt more capable AI agents, leaders need to be able to answer a practical question.
Your Agents Aren't the Only Agents to Worry About
As the section above described, healthcare organizations need to account for more than the AI agents they deploy. Agents can enter healthcare workflows through employees, software vendors, patients, caregivers, and partners. Each may represent a different person or organization, but healthcare organizations still need to know who is acting and what they are allowed to do.
For example, some agents support work from inside the organization. Healthcare teams may deploy them to help with prior authorization, claims, or care coordination. Employees may use personal productivity agents, while software vendors embed agents in applications the organization already uses.
Other agents approach from outside. A patient might use an agent to schedule an appointment, a member to check benefits, or a partner to exchange information. These agents create new ways to interact with healthcare services, including potential access to protected health information (PHI), clinical records, claims, benefits, and payment data.
Healthcare has long managed access for patients, clinicians, employees, partners, and applications. AI agents extend that challenge because they can use tools, move between systems, and take actions on someone’s behalf at machine speed and internet scale. An agent helping with an appointment, a claim, or care coordination may need access to sensitive clinical, benefits, or payment information to complete its task. Healthcare organizations need to recognize that agent activity and verify that each action stays within the authority it has been given.
The organization’s visibility into that activity can vary. An agent it deploys may have a defined owner and approved permissions. One introduced by an employee or embedded in a vendor’s application may be harder to identify and oversee. An agent representing a patient or partner may be managed entirely outside the organization.
Regardless of ownership, the healthcare organization first needs to recognize when an agent is interacting with its systems. It then needs to verify whom the agent represents, what that person or organization has authorized, and which information and actions the task requires. That visibility and control need to continue as the agent moves across systems.
How can healthcare organizations establish trust with every agent that reaches their systems, including those they didn’t build and don’t control?
An Identity for AI Framework for Each Type of Agent
Healthcare organizations need a consistent way to discover AI agents, establish their identities, and control access to systems and data. As agents take on more work, organizations need visibility into their activity, clear limits on their authority, and evidence of what they have done.
An identity framework connects those needs: discover, register, authenticate, authorize, and govern. Together, these capabilities help organizations verify agent identities, preserve whom they act for, and enforce access policies when they request a protected resource. Requests outside the agent’s approved permissions can be denied at that point, while governance provides oversight and an audit trail.
How organizations apply these capabilities depends on where an agent runs, who manages it, and whom it represents. Four types of healthcare AI agents illustrate those differences.
Four Types of Healthcare AI Agents
Employee agents
Employee agents are productivity, browser, and coding agents running on an employee’s device. In healthcare, they may help staff prepare documentation, find clinical guidelines, coordinate handoffs, or assemble prior authorization information.
Because their credentials can reside on the device, these agents can operate beyond the reach of traditional identity controls. Organizations need to discover them, understand their activity, and control their access to enterprise resources. Organizations also need a record of what these agents accessed and did, distinct from the employee’s own actions, to support investigations and audits.
For example: A referral coordinator uses a browser agent to gather documents for a specialist referral. The agent needs access to the relevant patient information, with limits that prevent it from retrieving unrelated records.
Consumer and partner agents
Consumer and partner agents act on someone’s behalf through digital channels. A patient might use one to schedule an appointment, a member to check a claim, or a partner to request an authorized exchange of records.
These agents create a new front door to healthcare services. Organizations need to verify whom they represent and enforce access policies that prevent them from doing more than that person or partner has authorized. They also need a record of what each agent accessed and did, and under whose authority, to support investigations and audits.
For example: A member authorizes a personal agent to check the status of an insurance claim. The payer needs to verify that relationship and allow the status check without granting permission to change coverage or payment details.
Workload agents
Workload agents are the organization’s digital workforce: assistants and autonomous agents running on self-hosted or third-party platforms. Healthcare examples include eligibility checks, claims processing, care-gap identification, and capacity management.
These agents can act faster than static access controls can accommodate. They need distinct identities, clear ownership, and limited permissions, with access policies enforced as they request resources. Human oversight helps maintain control over consequential actions. Organizations also need a record of the agent’s activity so they can review what it accessed and did, under whose authority, and with what approvals for audit purposes.
For example: A provider deploys an agent to assemble prior authorization packets. It retrieves the documentation needed for each request, while submission remains subject to review and approval by an authorized staff member.
SaaS-embedded agents
SaaS-embedded agents run inside a vendor’s application and may call back into the healthcare organization’s resources. They could operate within a contact center, revenue-cycle platform, or other healthcare application.
Although the vendor manages the agent, the healthcare organization needs visibility into what it can reach within its environment. Identity controls help enforce access policies that limit the agent to approved resources and actions. Organizations also need a record of what the agent accessed and did through the SaaS connection, under whose authority, and with what approvals, to support investigations and audits.
For example: An agent embedded in a contact center platform retrieves appointment and billing information to help resolve a patient inquiry. It should be able to retrieve the details needed for that inquiry without gaining permission to change billing information or access unrelated clinical records. Its activity should be logged so the organization can review what happened.
An Assigned Goal Is Not Permission to Access
Across all four agent types, the same distinction matters: what an agent is asked to accomplish and what it is authorized to do are separate things. In the Medicare incident discussed at the beginning of this blog, the agent was assigned to research public medicine spending but accessed a restricted government service without authorization.
In healthcare, those access limits are especially important because completing one task may involve several systems and different levels of permission. An agent may be allowed to retrieve information without changing it, prepare a request without submitting it, or assist a clinician without receiving all of that clinician’s access rights.
A valid login or token does not remove those limits. Authenticating an agent establishes its identity, while authorization determines what it can access and do. Access policies need to enforce its approved permissions as it requests resources or attempts protected actions.
The question healthcare leaders need to have the ability to answer at any point in time is: Is this agent, at this moment, authorized to take this action, and by whom?
Agentic AI in Healthcare Is Especially Consequential
Healthcare is critical infrastructure. Disruptions can interrupt essential services, delay treatment, and put patients at risk. Protected health information (PHI) is deeply sensitive, and healthcare organizations already face persistent targeting by cybercriminals. Introducing AI agents into this environment makes control over their access and actions especially consequential.
As agents work across clinical, coverage, and payment systems, their actions can affect both sensitive information and the delivery of care. Healthcare organizations need to address both the risks and the controls required in five areas:
- Patient safety: An inappropriate action in triage, referral routing, prior authorization, or care coordination can delay treatment or affect clinical outcomes. Organizations need to limit agents to approved tasks and require qualified human review for actions that could affect patient care.
- Regulatory accountability: Unauthorized access or incomplete audit records can expose organizations to compliance failures, investigations, and potential penalties. Organizations need to document which agent accessed what information, for what purpose, under whose authority, and with what approvals.
- Interoperability: As agents move information across EHRs, specialist systems, payer portals, and partner environments, they can expose sensitive data beyond its intended use or destination. Organizations need to enforce authorization for each access and transfer, preserving applicable consent and access restrictions across systems.
- Agent sprawl: Departments, employees, vendors, and partners can introduce agents faster than centralized governance can oversee them, leaving excessive permissions and risky activity undetected. Organizations need ongoing discovery, clear accountability, and regular reviews of agent access and activity.
- Human oversight: An agent acting beyond its authority can make a consequential decision without the right person’s knowledge or approval. Organizations need clear boundaries for independent action, approval requirements for sensitive decisions, and a way for authorized people to intervene.
Addressing these five areas depends on knowing what agents can access and enforcing limits on their activity. Yet research suggests those controls are often missing. In IBM’s 2026 Cost of a Data Breach study, 92% of organizations with AI-related breaches lacked proper AI access controls. Only 40% of the breached organizations studied reported using access controls on AI models and data.5
The study covered multiple industries, but the findings are particularly relevant wherever agents can reach protected health information. Least privilege—giving an identity only the access its task requires—needs to apply to agents as consistently as it does to people.
These requirements make identity a necessary foundation for responsible AI adoption. With visibility, limited permissions, and accountable oversight, healthcare organizations can put agents to work while maintaining control over their access and activity.
The Journey to Trusted Identity for AI Agents
Strengthening access controls for a new class of user, AI agents, starts with applying familiar identity principles. Healthcare leaders need to know where agents are operating, who is responsible for them, what they can access, and what they have done. Five steps connect that visibility to control.
- Discover: Where are agents operating?
Identify agents deployed across—and interacting with—the organization’s environment, including employee devices, cloud platforms, SaaS applications, and patient and partner channels. Discovery helps reveal activity that might otherwise escape oversight. - Register: Which agents are we responsible for?
Give agents the organization directly manages distinct identities and associate them with accountable owners. For agents managed by vendors or external parties, establish the identity and ownership information needed to make informed access decisions. The approach may differ, but accountability remains essential. - Authenticate: Can we verify the agent’s identity?
Verify which agent is requesting access and preserve the identity of the person or process it represents. A prior authorization agent should be recognized as an agent acting for an authorized clinician or team, with that relationship kept clear throughout the interaction. - Authorize: What is the agent allowed to do?
Enforce permissions and access policies as the agent requests resources or attempts protected actions. Limit access to what its task requires, and require human approval where appropriate. For example, permission to prepare a prior authorization packet should not automatically include permission to submit it. - Govern: What has the agent done, and is its access still appropriate?
Collect activity records, review access, and maintain accountable ownership as agents and their responsibilities change. Organizations need evidence to investigate unusual activity, support audits, and determine whether permissions should be retained, adjusted, or removed.
Together, these steps help healthcare organizations give agents useful access while keeping their authority limited and their activity accountable. And timing is everything. Enter runtime identity.
Runtime Identity Moves Trust to the Moment of Action
The identity journey establishes who an agent is and what it is allowed to do. Runtime identity brings that information into access decisions as the agent works. When it requests a protected resource or attempts an action, the organization can evaluate whether that request is authorized in its current context.
That decision needs to account for:
- Which agent is acting and whom it represents
- What it is trying to do and which information or system is involved
- Whether its delegated permissions cover the request
- Whether current conditions or risk warrant additional checks
- Whether an authorized person needs to approve the action
Authentication verifies the agent’s identity. Authorization determines whether the requested access or action is permitted. Runtime control lets an organization evaluate the action in its current context and prevent access outside the approved scope in real-time (runtime). Enforcing policies in the real-time point of decision/action allows organizations to deny requests outside the agent’s approved scope, even after it has successfully authenticated.
For example, an agent preparing a prior authorization packet requests access to a patient’s documentation. At that moment, identity controls verify the agent, whom it represents, and whether its permissions cover the requested records. If it later attempts to submit the packet, that action requires a separate authorization decision, including any required human approval. Logging these requests and authorization decisions creates an audit trail showing which agent acted, under whose authority, and what was permitted or denied.
Government guidance reinforces this approach. On April 30, 2026, the NSA announced the release of “Careful Adoption of Agentic AI Services”, jointly authored by cybersecurity agencies from Australia, Canada, New Zealand, the United Kingdom, and the United States. The guidance recommends continuously verifying identity and authorization at runtime, using a centralized policy decision point for each request.6˒7
Delegation Keeps the Human and the AI Agent Distinct
Runtime access decisions depend on knowing both which agent is acting and whom it represents. That distinction becomes harder to maintain when an agent uses a person’s credentials or broadly inherits their permissions. The resulting activity may appear to come from the person, obscuring what the agent did and whether it had permission.
Delegated authority makes that relationship explicit. The agent acts with its own identity and a defined scope of authority granted on someone’s behalf. A clinician can authorize an agent to gather documentation for a referral, for example, without giving it every permission the clinician holds.
Keeping those identities separate helps organizations answer essential questions:
- Did the person or the agent take the action?
- What had the agent been authorized to do?
- Did the action stay within that authority?
- Who was responsible, and was human approval required?
An agent should act on someone’s behalf without becoming indistinguishable from that person.
Preserving that distinction supports limited access, meaningful oversight, and an audit trail that connects each agent’s actions to the authority behind them.
Identity Bridges Healthcare AI Risk and Opportunity
Realizing the value of AI agents requires confidence that they can reach the resources they need while staying within the authority they have been given. With that foundation, healthcare teams can use agents to enable better care, reduce administrative work, and drive revenue.
Identity provides the foundation. Discovery makes agent activity visible. Authentication and delegation establish which agent is acting and whom it represents. Runtime identity brings that context into access decisions at the moment an agent requests a protected resource or attempts an action, allowing authorization policies to enforce its approved permissions. Governance provides oversight and evidence for review. Together, these capabilities extend healthcare’s identity practices to a new class of digital participant.
The Medicare incident illustrates why a useful assignment alone cannot define an agent’s authority. Healthcare organizations will not control every agent that approaches their systems, but they can enforce access policies at protected resources and require approval for consequential actions.
The path forward is to make identity part of AI adoption from the start: know which agents are interacting with the organization, define what they are allowed to do, and maintain the controls and records needed to hold their activity accountable.
With identity as the foundation, healthcare organizations can put AI agents to work reducing administrative burden, accelerating patient and member services, and freeing staff to focus on care, while protecting sensitive information and maintaining accountability.
Frequently Asked Questions
- ABC News - OpenAI Hacked Medicare Portal, Prime Minister Anthony Albanese Says
- OpenAI - How We Will Do Better for Australia
- OpenAI - The Hugging Face Incident and the Road Ahead
- HIPAA Journal - IBM Cost of a Data Breach Study 2026
- IBM - AI-Powered Adversaries and the Enterprise Risk Challenge: Preparing for the New Reality
- National Security Agency - NSA Joins the ASD's ACSC and Others to Release Guidance on Agentic Artificial Intelligence
- Australian Signals Directorate - Careful Adoption of Agentic AI Services