Weill Cornell Medicine patient notice issued the week before November 12, 2026; incident discovered September 6, 2026; underlying research activity June 2026
Patient Data Into an Unapproved AI Tool: Weill Cornell Notifies Patients in the Second Shadow AI Incident in Weeks
Weill Cornell Medicine is informing patients that information from a radiology research project was entered into artificial intelligence tools without authorization, in what coverage describes as the second shadow AI data incident at a major academic medical center in the past few weeks. The health system said that during a radiology research study this past June, information from Weill Cornell Imaging used at NewYork-Presbyterian was entered into AI tools and processed using machine learning without authorization and in violation of its own policies. Weill Cornell discovered the incident on September 6 and immediately initiated an investigation; on discovery it suspended the research activity, requested that all involved data be deleted, and provided additional education to reinforce its requirements for appropriate AI tool use. The data involved included medical record numbers and radiology accession numbers, dates and times of radiology services, the place care was received, and clinical information relating to radiological examinations, such as diagnoses, treatment, surgical and procedural histories, and medication information. The notice states what the data did not include: no patient names, addresses, telephone numbers, email addresses, Social Security numbers, or patient-specific financial information. Weill Cornell said the data used was limited, that there is no indication any patient information has been misused, and that it has so far found no evidence that unauthorized individuals accessed the protected health information. The earlier incident, reported in October, was at Johns Hopkins Bayview Medical Center in Baltimore, where a staff member reportedly uploaded a spreadsheet to a public AI website that had not been approved for use, exposing data of more than 3,000 patients.
Executives
Two academic medical centers in a matter of weeks, same mechanism: a staff member with legitimate access used an unapproved AI tool for real work, and the organization learned afterward. That is a governance gap, not a hacking story, and it is the pattern the acceptable use conversation exists to prevent. The questions are standing ones: is there an approved AI tool that does the job staff are reaching for, do staff know which tools are approved and for what data, and would your logging show patient or customer data entering an unapproved service today?
CIO / CISO
Treat shadow AI as a detection problem as much as a policy problem. Inventory the AI services reachable from clinical and research networks, log uploads to unapproved services where the tooling allows, and give research teams an approved path for exactly the workload in this notice, imaging-adjacent analysis, so the compliant route is also the easy route. When an incident occurs, the Weill Cornell sequence is the template: suspend the activity, request deletion from the tool provider in writing, document the data elements involved and the ones excluded, and notify on the evidence.
Engineers
Support the program with technical boundaries: egress controls and data loss prevention rules tuned for uploads to consumer AI services, separate research environments with named, approved models, and de-identification tooling for imaging data before any analysis step. Medical record numbers and accession numbers are identifiers even when names are absent, so pipelines should treat them as protected fields in schemas, logs and exports, not as harmless metadata.
Everyone (non-IT)
The rule for any workplace AI tool is simple: if the tool has not been approved for the data in front of you, the data does not go in, however helpful the tool looks and however ordinary the task feels. That applies to a spreadsheet, a scan, a transcript or a screenshot. If you are unsure whether a tool is approved for patient, customer or internal information, ask before pasting; the two incidents in this story both started with someone doing their job quickly, not with anyone intending harm.