Johns Hopkins Bayview Medical Center began mailing letters on October 22 to 3,182 patients whose records were pasted into a consumer AI chatbot by a staff member, in one of the first U.S. hospital disclosures of its kind.
The upload was caught the same day, the Baltimore hospital said. On September 21, an employee dropped a spreadsheet of patient information into an AI platform the health system had never reviewed or approved for clinical use. The file listed men and women seen at the hospital at some point from spring through September 2026.
Fields at risk ran from names and treatment dates to record and account numbers, medications, the treating department and physician names. Social Security numbers and financial accounts were not present, the hospital said.
Cleanup followed in sequence. From the tool’s history the worker removed the prompt, thread and document, and Johns Hopkins asked the vendor to wipe the data too. Access is now limited to approved applications, with retraining across the workforce.
Patients are being offered 12 months of credit monitoring through Experian IdentityWorks, and a toll-free call center is taking their questions.
For hospital security chiefs, the case shows the weakest link in AI governance is a browser tab. Shadow AI use rarely trips an intrusion detection system and leaves no log in the EHR. Block unapproved domains at the gateway, inventory sanctioned tools, and train staff that public chatbots retain what is typed into them.
Regulators treat an upload to an unapproved tool as a disclosure of PHI, not an internal mishap. Johns Hopkins Medicine has been expanding its own AI work, including a partnership announced this month with OpenEvidence, which makes clear approved and unapproved tools now sit side by side on clinical desktops.
