Leveraging LLMs for Policy, Regulatory, and Compliance in IoMT - Opportunities, Risks and Safeguards

The talk opens with the state of healthcare breaches and the connected medical device attack surface in hospitals. It points to flat networks, overprivileged identities, and devices that have stayed unpatched for years, with many connected devices per bed. Availability and integrity matter because a device cannot simply be taken offline to patch it. The team's threat model showed that a poisoned retrieval knowledge base could steer a clinical AI agent into suppressing alerts or changing dosing.

The middle section reviews existing risk frameworks and finds little fit for clinical settings. The team then presents LLM-based approaches. One uses graph-based access control (NGAC) to limit what AI agents can reach, so that injected instructions cannot widen their privileges. Another is a knowledge-graph tool that compares vendor policies against HIPAA-like and NIST guidance and produces gap reports. Data security, covering encryption in transit and at rest, was described as work still in progress.

The closing part covers risks: prompt injection, hallucinated assurance, and reliance on external or cloud-hosted models. The advice is to keep humans in the loop, treat model output as untrusted input, and layer defenses rather than rely on one product. In the Q&A, the presenter explained that linking specific devices to specific control gaps can turn manual compliance review into automated risk assessment.