Weill Cornell Medicine has begun notifying patients that protected health information tied to a radiology research study was pushed into AI and machine learning tools without authorization, a violation of its own policies. The New York City medical school and biomedical research institution discovered the problem on September 6, 2026, and traced the uploads to the window between June 10 and June 20.
Numbers without names, but still PHI
The exposed records came from imaging patients at Weill Cornell Imaging at NewYork-Presbyterian and NewYork-Presbyterian Hospital. Names, contact details, Social Security numbers and financial data were not included, according to the notice, but the uploads carried medical record numbers, radiology accession numbers, exam dates and times, facility and provider names, and clinical content covering diagnoses, treatments, surgical history and medications. Under HIPAA that combination is identifiable without a name attached. The organization found no misuse, has suspended the research, asked for deletion and retrained the personnel involved. It has not disclosed how many patients were affected.
The quiet channel into consumer models
The same week brought a second Washington state case: a staff member at Lake Burien Transitional Care Facility drafted client case summaries in a personal ChatGPT account, covering names, diagnoses and medications for 12 patients, with four more Western State Hospital patients caught in a similar pattern. A Wolters Kluwer Health survey found 40% of hospital workers knew of colleagues using unapproved AI tools.
Controls that do not depend on intent
Shadow AI in clinical research is an access-governance problem, not a training problem. Teams running imaging studies should inventory which models touch study data, restrict research datasets to approved environments, log every export of accession numbers, and require a named data steward before any assistant is trialed. Deletion requests to consumer vendors come with no receipt.
