A review published in Nature on August 19 systematically maps the security and safety risks of large language models in healthcare, arguing that clinical adoption is racing ahead of risk assessment. The authors sweep the literature on LLM security, from the models themselves to their integration with hospital workflows and interactions with clinicians.
The review organizes hazards by development stage, covering design, data, model, inference and environment, and lays out protective layers that run from core optimization objectives and knowledge integrity to alignment and human interaction. It also classifies threats by how clinically relevant they are today.
The authors end with a perspective on mitigation, spelling out which responsibilities fall to developers, health systems and regulators as LLMs move into documentation, triage and patient messaging.
Why it matters: hospitals are deploying LLMs with little shared vocabulary for evaluating safety. A single framework that maps where attacks and failures can occur gives buyers a checklist before rollout, and gives vendors a roadmap for hardening their models.
