Patient Data Exposure Through AI Tools: The Johns Hopkins Case
Artificial intelligence is transforming healthcare, helping professionals summarize information, streamline administrative work, and improve productivity. But as AI adoption accelerates, a critical question remains: What happens when employees share sensitive information with AI tools that their organizations have not approved?
A recent incident at Johns Hopkins Bayview Medical Center in Baltimore, Maryland, highlights the risks.
In September 2026, a staff member uploaded a spreadsheet containing patient information to a public AI tool that had not been approved for that purpose. The incident led the medical center to notify approximately 3,182 patients that their information may have been exposed.
The incident is a reminder that AI data security is not simply a technology issue. It also requires clear policies, visibility into AI usage, and safeguards that help prevent sensitive information from leaving an organization’s control.
What Happened at Johns Hopkins Bayview Medical Center?
According to reports published in October 2026, Johns Hopkins Bayview Medical Center discovered the incident on September 21, 2026, during an internal review. On the same day, a staff member mistakenly uploaded a spreadsheet containing patient information to an external AI platform that the hospital had not approved for this use.
The spreadsheet contained information associated with patients admitted and treated at the hospital between May and September 2026.
The information that may have been exposed included:
- Patient names
- Medical record numbers
- Account numbers
- Departments providing treatment
- Medications administered
- Healthcare provider names
Following the discovery, the medical center initiated an investigation, directed the employee to delete the prompt, conversation thread, and document from the AI tool’s history, and contacted the AI provider to request deletion of the information.
The hospital also began notifying affected patients and reinforcing its policies and training concerning protected health information and unauthorized applications.
The central concern is straightforward: sensitive information can leave an organization’s intended security boundaries through an everyday workplace action.
Source: Hoodline’s report on the Johns Hopkins Bayview incident: https://hoodline.com/2026/10/johns-hopkins-bayview-warns-3-182-patients-after-staffer-fed-data-to-ai-tool/
Why Patient Data Exposure Through AI Tools Is a Growing Concern
Employees increasingly turn to generative AI to simplify routine tasks. They may ask an AI assistant to summarize documents, analyze spreadsheets, draft reports, or organize information.
These activities can improve productivity. However, the risk changes when employees submit confidential business information, patient records, or other regulated data to tools that have not been assessed and approved by their organizations.
1. Sensitive Information Can Be Shared Accidentally
Employees do not always recognize that a file contains sensitive information or understand the implications of uploading it to an external AI platform.
A spreadsheet submitted for analysis may contain patient identifiers, treatment information, account details, or other confidential records. Even when the employee’s intention is legitimate, the information may be disclosed to a service that the organization has not authorized to process it.
This is why security controls should not rely entirely on employees remembering every policy.
2. Unauthorized AI Usage Can Be Difficult to Detect
Employees may access AI tools through personal accounts, browser-based applications, or services that have not been formally reviewed by IT and security teams.
This behavior is often associated with Shadow AI: the use of AI applications without an organization’s knowledge, approval, or appropriate oversight.
Without adequate visibility, security teams may struggle to determine which tools employees use, what types of information they share, and whether those activities comply with organizational policies.
3. Deleting a Conversation May Not Resolve Every Risk
Deleting a prompt or uploaded file from an application’s visible history does not, by itself, establish what happened to the information across the provider’s systems.
Data retention, deletion processes, access controls, and other handling practices depend on the particular service, account configuration, and applicable terms.
Organizations should therefore investigate the scope of a disclosure, follow up with the relevant provider, and verify deletion or remediation where possible rather than assuming that removing a conversation resolves the incident.
4. Healthcare Information Requires Stronger Safeguards
Healthcare organizations handle information that can reveal highly personal details about individuals, including their medical histories, treatments, medications, and relationships with healthcare providers.
In the United States, the Health Insurance Portability and Accountability Act (HIPAA) establishes requirements for protecting applicable protected health information. Whether a particular AI workflow satisfies HIPAA requirements depends on the circumstances, including the information involved, the parties processing it, and the safeguards and contractual arrangements in place.
Using an AI tool does not automatically make a workflow noncompliant. However, organizations must evaluate whether the tool and its intended use meet their legal, privacy, and security obligations.
The Connection Between Shadow AI and Data Security
The Johns Hopkins Bayview incident illustrates how an organization can face data exposure even when an employee is not deliberately trying to bypass security.
A worker may choose an AI tool because it is convenient, familiar, or capable of completing a task quickly. If that tool has not been approved, the organization may lack the necessary controls to govern the interaction.
This creates a gap between AI adoption and AI governance.
Effective AI governance should answer several questions:
- Which AI platforms are employees permitted to use?
- What types of data can be shared with each platform?
- Can sensitive information be detected before it leaves the organization?
- Can employees be warned or blocked when a prompt violates policy?
- Can security teams investigate AI-related activity when an incident occurs?
- Are approved alternatives available so employees can complete legitimate tasks securely?
Organizations need a way to turn AI security policies into controls that operate during everyday AI interactions.
How Organizations Can Reduce the Risk of Sensitive Data Exposure Through AI
Preventing accidental disclosures requires a combination of governance, employee education, and technical controls.
1. Establish an Approved AI Usage Policy
Organizations should maintain a clear inventory of approved AI applications and explain which tasks and data types are permitted for each.
The policy should address public AI tools, personal accounts, confidential documents, regulated information, and exceptions requiring formal approval.
Employees should also understand where to go when they need an AI capability that is not currently approved.
2. Improve Visibility Into AI Usage
Security teams need appropriate visibility into how employees access AI applications and whether their activity complies with organizational policy.
Monitoring can help organizations identify unauthorized applications, understand emerging usage patterns, and investigate potentially risky interactions.
Visibility should be implemented with suitable privacy protections and access restrictions.
3. Detect and Protect Sensitive Information Before Submission
One of the most important safeguards is to assess the information being shared before it reaches an external AI service.
Depending on the workflow, technical controls can identify sensitive data, warn users, block prohibited submissions, or redact information that does not need to be disclosed.
For example, a document may contain information required to complete an administrative task alongside patient identifiers that are unnecessary for that task. A properly designed workflow can help protect those identifiers while preserving the information needed for the approved use.
4. Enforce Policies Across Supported AI Workflows
Security policies should be applied consistently across the AI applications and workflows an organization supports.
Controls may include restricting unauthorized applications, allowing only approved platforms, preventing prohibited data transfers, and guiding employees toward compliant alternatives. Coverage will depend on the technologies deployed, their configurations, and the organization’s environment.
5. Maintain Auditability and Incident Response Readiness
Organizations should be able to investigate suspicious activity, understand which policies were applied, and document relevant events.
Audit records can support incident response, internal reviews, and compliance processes. Access to these records should be appropriately restricted, and retention should follow organizational requirements.
A defined response process should also cover containment, investigation, provider coordination, required notifications, and corrective action.
How SecurePledge Can Help Organizations Strengthen AI Data Security
The Johns Hopkins Bayview incident demonstrates why organizations need more than a policy telling employees not to share sensitive information with unapproved AI tools.
They need safeguards that help enforce those policies in practice.
SecurePledge is designed to help organizations protect sensitive information across AI interactions and digital workflows through capabilities for data protection, AI usage controls, and governance.
SecurePledge is designed to help organizations protect sensitive information across AI interactions and digital workflows through capabilities for data protection, AI usage controls, and governance.
Redactor: Protect Sensitive Information
SecurePledge Redactor is designed to identify and redact sensitive information in supported content before it is shared with an AI service.
This can help reduce the risk of exposing personal or regulated information when employees need AI assistance with documents or other work materials.
For healthcare organizations, the goal is to reduce unnecessary disclosure while enabling appropriate, approved AI use.
Protector: Support Safer AI Usage
SecurePledge Protector helps organizations apply controls to supported browser-based AI workflows.
These controls can help identify unauthorized AI usage, warn or block users when prompts contain sensitive information, and guide employees toward approved AI platforms, depending on configuration and supported coverage.
This approach helps address a key Shadow AI challenge: employees may use external AI services without understanding the data security implications.
Governor: Support Governance and Accountability
SecurePledge Governor provides a governance layer intended to support consistent policy enforcement and auditability across the platform.
Governance capabilities help organizations connect their AI security policies with the controls and oversight needed to manage AI usage more systematically.
Together, these capabilities support a proactive approach to AI security: protect sensitive information, govern access to AI tools, and improve oversight of AI interactions.
For organizations handling patient information or other confidential data, the appropriate implementation should be evaluated against their specific workflows, regulatory obligations, and security requirements.
The Bigger Lesson: AI Adoption Must Be Matched by AI Governance
The Johns Hopkins Bayview incident is a reminder that data exposure does not always begin with a sophisticated cyberattack. Sometimes, it begins when an employee uploads a spreadsheet to a tool that appears useful for an everyday task.
As organizations expand their use of generative AI, the challenge is to enable productivity without losing control over sensitive information.
That requires more than awareness campaigns or written policies. Organizations should combine approved AI workflows, employee training, visibility into AI usage, preventive data protection, and reliable incident response procedures.
AI adoption should move forward with security built into the workflow, not added after an incident.
SecurePledge helps organizations take a more controlled approach to AI interactions by supporting sensitive data protection, AI usage controls, and governance across supported workflows.
Learn more about SecurePledge
Protect Sensitive Data Before It Reaches AI
Uncontrolled AI usage can expose sensitive information and create unnecessary security risks. SecurePledge helps organizations protect sensitive data, govern AI usage, and implement security controls across supported AI workflows.
Explore SecurePledge to learn how your organization can adopt AI with greater confidence.
Sources and References
- Hoodline, Johns Hopkins Bayview warns 3,182 patients after staffer fed data to AI tool: https://hoodline.com/2026/10/johns-hopkins-bayview-warns-3-182-patients-after-staffer-fed-data-to-ai-tool/
- Becker’s Hospital Review, Johns Hopkins alerts 3,182 patients after worker uploads data to unapproved AI tool: https://www.beckershospitalreview.com/healthcare-information-technology/ai/johns-hopkins-alerts-3182-patients-after-worker-uploads-data-to-unapproved-ai-tool/
- CBS Baltimore, Johns Hopkins Bayview patient data incident coverage: https://www.cbsnews.com/baltimore/news/johns-hopkins-bayview-patients-data-breach-3100-people/