Article
What happened at Bayview
On September 21, 2026, during an internal review, a Johns Hopkins Bayview Medical Center staff member uploaded a spreadsheet to an outside AI platform the hospital had not approved. The file covered 3,182 patients treated from May through September 2026 and contained names, admission and treatment dates, medical record and account numbers, medications, treating departments and provider names. The hospital says Social Security numbers and financial account data were not included.
Bayview caught the upload the same day. It had the employee delete the prompt, chat thread and document, asked the tool’s operator to purge the data, and began mailing patient letters on October 22 with 12 months of free credit monitoring. It also says it is tightening policy and retraining staff on which tools are allowed.
There was no attacker here. A staff member was trying to work faster with a tool that was easy to reach. That is the shadow AI problem in miniature, and most organizations cannot say how often it occurs.
Federal guardrails are thinning
On October 26, 2026, ONC released the HTI-5 final rule, published in the Federal Register two days later. It cuts 14 of 60 health IT certification criteria and ends the requirement that certified EHRs support source attributes and risk management for predictive AI, the so-called model cards. The first changes take effect December 28, 2026. Providers still carry their federal duty under 45 C.F.R. § 92.210 to identify and reduce bias in predictive tools. In practice, transparency about the AI inside your EHR now depends on your contracts and your own oversight.
Peers are working on the same gap: on October 29, 2026, the AHA’s West Health Accelerator hosted a virtual session on how hospitals evaluate AI products and set guardrails.
What executive-level governance should include
- A clear policy. Define AI broadly enough to cover public chatbots, browser add-ons and ambient documentation tools. Bar protected health information from any tool not on the approved list. Name an accountable executive and a cross-functional committee spanning clinical, compliance, security, legal and operations, with regular reporting to the board.
- An approved-tool list people will use. If staff have no sanctioned, contract-covered option for summarizing or drafting, they will find an unsanctioned one. Publish the list, offer a fast intake path for new requests, and require vendor terms on data retention, use of your data for model training, and the transparency details HTI-5 no longer guarantees.
- Detection. Pair the policy with network and data loss prevention controls that flag patient data headed to public AI sites. Policy without visibility is a hope, not a control.
- An AI-specific incident playbook. Extend the breach plan to cover deletion inside the tool, written purge requests to the vendor, privacy risk assessment and the notification decision. Bayview’s same-day discovery is the standard to test against.
Sources
- Johns Hopkins alerts 3,000+ patients after data sent to AI tool (WMAR 2 News, Oct 23, 2026)
- Johns Hopkins data exposure caused by employee using unauthorized AI tool (The Baltimore Banner, Oct 25, 2026)
- HHS Finalizes HTI-5, Narrowing Health IT Certification and Deferring Information Blocking Changes (McGuireWoods, Oct 30, 2026)
- Virtual session to discuss how hospitals put AI governance into practice (AHA News, Oct 26, 2026)

