AI tool leads to patient data breach at hospital

A single spreadsheet, a consumer AI tool, and 3,182 patient notification letters—these are the key elements of a September 21 incident at Johns Hopkins Bayview Medical Center in Baltimore. This event highlights the significant consequences for healthcare organizations when unauthorized AI tools are used, leading to potential data breaches.

Hospital’s AI Oversight Leads to Data Exposure

According to WMAR-2 News, an employee at the hospital uploaded a spreadsheet containing patient data to an unapproved public AI tool. The spreadsheet included information on patients treated between May and September, such as names, admission and treatment dates, medical record numbers, account numbers, treating departments, medications, and provider names. Notably, Social Security numbers and financial account details were not part of the exposed data, as confirmed by the hospital.

The upload was detected the same day through an internal review. The employee promptly removed the prompt, thread, and document from the AI tool. Johns Hopkins then requested the AI company to delete the data and implement additional safeguards. On October 22, the hospital began sending notification letters to affected patients, offering 12 months of free credit monitoring.

Straight Arrow News reported that the hospital has not disclosed the name of the AI tool or confirmed whether the vendor successfully erased the data. While deleting a chat removes it from the user’s history, it does not guarantee removal from the company’s servers.

In a statement, Johns Hopkins emphasized its commitment to protecting patient privacy and security. The hospital also noted it is enhancing policies and staff training regarding the use of approved AI tools, as reported by The Baltimore Banner.

Regulatory Challenges for Smaller Healthcare Practices

Even smaller practices, such as those with five physicians, face the same stringent regulations as larger institutions. Under the HIPAA Breach Notification Rule, any unauthorized disclosure of patient data is presumed a breach unless a risk assessment demonstrates a low likelihood of compromise. This assessment evaluates four factors: the type of data involved, who received it, whether it was accessed or viewed, and the extent of risk mitigation. Uploading data to a consumer AI vendor with unclear data retention practices makes it difficult to argue against breach criteria.

If a breach occurs, the practice must notify affected patients in writing within 60 days of discovery. Breaches involving 500 or more individuals also require notification to the Department of Health and Human Services (HHS) and, if applicable, prominent media outlets in the affected state. Smaller breaches must be logged and reported to HHS within 60 days after the end of the calendar year. State laws may impose additional deadlines and requirements, such as notifying the attorney general.

Notification is only part of the issue. HHS guidelines on cloud services classify any service handling electronic protected health information as a business associate, even if it cannot view the data. Using such a service without a business associate agreement violates HIPAA. Personal chatbot accounts used by employees lack these agreements, further complicating compliance.

Developing Effective AI Policies

The Johns Hopkins incident is not isolated. A December 2025 Wolters Kluwer Health survey of 518 hospital and health system providers and administrators found that 17% admitted to using unauthorized AI tools at work. Among those, 45% cited faster workflow as the primary reason.

Governance has struggled to keep up. Only 27% of physicians and nurses reported awareness of formal AI governance policies at their organizations, up from 21% the previous year, according to Wolters Kluwer’s 2026 Future Ready Healthcare survey conducted by Ipsos in March among 203 physicians and 152 nurses. Medical Economics recently highlighted that staff often resort to external tools when approved AI systems reach usage limits.

With notification letters sent, the 60-day federal deadline for a breach of this size expires in late November.

Essential Elements of an AI Policy

A written policy that holds up after an incident should cover a few basics: a named list of approved tools, staff not having to guess which are permitted, every approved tool requiring a signed business associate agreement, and a flat ban on patient data in personal accounts, browser extensions, and free tiers of paid tools. Technical blocks on consumer AI sites in practice devices and networks, along with an immediate reporting duty, are critical. Documented training with sign-off records demonstrates the policy existed before any mistake occurred.