As artificial intelligence (AI) systems increasingly become integral to healthcare operations, organizations face a critical challenge: how to incorporate AI into their HIPAA risk analysis frameworks responsibly and effectively. Currently, there is no explicit regulatory guidance on AI in healthcare under HIPAA. This gap leaves cybersecurity and compliance professionals navigating uncharted territory, tasked with safeguarding electronic protected health information (ePHI) while leveraging AI’s potential.

Core Elements of HIPAA Risk Analysis Remain Foundational

At its core, HIPAA risk analysis involves several fundamental components:

• Inventory of Organization: Understanding the organizational units and stakeholders involved in handling ePHI.
• Inventory of Data: Cataloging all ePHI data types, sources, and flows.
• Inventory of Systems: Identifying all IT systems and applications that create, receive, maintain, or transmit ePHI.
• Macro/External View of Threats: Recognizing external and internal threat actors and vectors.
• Controls Assessment: Evaluating existing safeguards and security controls.
• Qualitative Review: Synthesizing findings to prioritize risks and mitigation efforts.

These foundational elements remain essential when integrating AI, but must be adapted to address AI’s unique characteristics and risks.

Expanding the HIPAA Risk Analysis Scope

AI Systems and Agentic Subsystems

A critical first step is tagging any system or subsystem that interacts with ePHI. AI components can be embedded deeply within applications or workflows, often operating autonomously or semi-autonomously. It is prudent to:

• Break out agentic subsystems within applications – these are AI modules or “agents” performing distinct functions.
• Enumerate all agents and associated tools, detailing their roles, access methods, and data interactions.

This granular inventory enables targeted risk analysis and control assessment.

Detailed Access and Workflow Analysis

For each AI agent or subsystem, document:

• Access Methods: Are interactions automated workflows, conversational interfaces, or hybrid?
• Logging Practices: Create assurance that AI interactions involving ePHI are logged at the appropriate level of detail, with clear attribution of access, whether human or non-human agents acting as brokers of ePHI.
• Tools and Interfaces: Catalog all AI runtimes and tools sanctioned by the organization (e.g., Cowork, control plane AI services), and monitor for unapproved tool use.

Threat Modeling and Risk Assessment of AI Agents

AI introduces novel threats such as adversarial attacks, data poisoning, and model manipulation. Therefore:

• Conduct a threat model and risk assessment for each AI agent, considering both technical vulnerabilities and operational risks.
• Implement change control and reassessment processes for modifications to AI models, tools, instructions, and user interfaces to maintain security posture and compliance over time.

Addressing Training and Model Management Risks

If the organization performs AI model training or fine-tuning:

• Evaluate data quality and bias risks rigorously, as biased or poor-quality training data can lead to discriminatory or erroneous outcomes impacting patient safety and compliance.
• Document controls and validation processes to mitigate these risks.

Managing Unapproved AI Use

Given the rapid proliferation of AI tools, users may sometimes employ unapproved AI services. Organizations should:

• Acknowledge the risk of unapproved AI use.
• Document training, preventive, and detective controls to mitigate insider threats and accidental exposures.

Logging and Accountability: The Emerging Role of Tracing-Level Logs

Logging is paramount for HIPAA compliance and forensic readiness. As AI systems grow more complex, traditional logging may not provide sufficient granularity to fully understand AI behavior and data flows. This is where tracing-level logs become crucial.

Tracing-level logs capture detailed, fine-grained information about the execution path of requests and data through distributed systems and AI agents. They record sequences of operations, timing, metadata about components involved, and contextual data that help reconstruct how ePHI is accessed, processed, or transmitted by AI subsystems.

These logs are often generated and managed using open standards like OpenTelemetry, a widely adopted observability framework. OpenTelemetry provides instrumentation libraries and exporters that collect traces, metrics, and logs from applications and services, then ship them to monitoring and analysis platforms. This capability allows organizations to:

• Gain comprehensive visibility into AI workflows and data handling.
• Detect anomalies or unauthorized access in near real-time.
• Support detailed audit trails required for HIPAA Privacy and Security Rule compliance.

Given the evolving regulatory landscape and increasing scrutiny on AI transparency, collecting and maintaining tracing-level logs may become a required aspect of HIPAA-compliant logging practices in the future.

Conclusion: Applying Lessons Learned to New Frontiers

While explicit HIPAA guidance on AI remains forthcoming, organizations must act proactively. Organizations need to adapt their established risk analysis principles: inventorying systems and data, modeling threats, assessing controls, and maintaining rigorous logging and including tracing-level logs.  In doing so, healthcare entities can responsibly integrate AI technologies without compromising compliance or security.

This approach fosters a risk-aware culture that embraces innovation while protecting patient data, setting a strong foundation for future regulatory developments.  It also demonstrates proactive due-diligence to the OCR by attempting to address emerging technologies before any specific guidance has been issued.