AI security in healthcare: what the information security standard asks of an AI environment
Short answer
Dutch healthcare organisations must run their information security under NEN 7510. The December 2024 version builds on ISO 27001:2022 and asks for risk-based choices; certified organisations must have moved over before 20 February 2027. AI brings new risks into that system: an extra processor, output that lands in the record, broader access to data and logging that must match the access-logging standard. An AI environment in healthcare therefore belongs inside the management system, not beside it.
The standard in brief
NEN 7510 is the Dutch standard for information security in healthcare. Part 1 describes the management system, part 2 the controls. The new version appeared on 16 December 2024, based on ISO 27001:2022 and ISO 27002:2022, with fewer prescriptive controls and more emphasis on the organisation’s own risk assessment. Organisations with a certificate must complete the transition before 20 February 2027. The healthcare inspectorate checks against it, and NEN 7512 (secure exchange) and NEN 7513 (logging of access to the record) go with it. Hospitals elsewhere work under ISO 27001 and national rules of the same shape.
The standard does not mention AI. But every AI application that processes patient data falls under it, like any other system. See AI in healthcare: patient data, NEN 7510 and the GDPR for the privacy side; this article is about security.
What AI adds to the risk register
| Risk | Where it comes from | Control |
|---|---|---|
| A processor out of sight | A staff member uses a public AI service for a letter or summary | Policy, training and an environment of your own that works better than the public one |
| Output in the record | An AI summary or report is adopted without checking | Review by the clinician, source references, marking of AI output |
| Access too broad | An assistant searches more records than the user may see | Access per role following the EHR; see Access control for AI |
| No logging | Nobody can see afterwards who had which record read by AI | Audit log per question and answer, matching the access-logging standard |
| Manipulation | A document or email carries hidden instructions for the model | AI only reads and proposes, never acts on its own |
| Supplier as chain | The model runs at a provider, sometimes outside the EU | A node that belongs to the institution, or a dedicated node in a Dutch datacentre |
Requirements for an AI environment in healthcare
- Inside the management system. The AI environment is in the risk register, has an owner and is covered by the internal audit.
- Access through the existing identity. Sign-in with the same accounts and roles as the EHR, no separate passwords.
- Logging that can be traced. Who, when, which record, which question. Retention per your own policy, exportable to the SIEM.
- No training on your data. Demonstrably: with open-weight models on your own hardware it is a fact, with a service it is a contract clause you must check.
- Continuity. What happens if the connection or the supplier drops out? A node in your own building keeps working.
- Incident process. A wrong answer that lands in a record is an incident. See Step-by-step plan for a data breach with AI.
How to show it
The inspectorate and a certification body want evidence, not promises: the risk register with the AI application in it, the processing agreement or the own hardware, the role matrix, an export of the audit log and the evaluation of the pilot. Record for every AI application which data goes in, who checks the output and when the application is reassessed. That is the same file the AI Act will ask for, so you do the work once.
How werqly fits in
werqly runs on the healthcare organisation’s own node, in its own server room or as a dedicated node in a Dutch datacentre, with open-weight models that do not learn from your data. werqly AI Control supplies the roles, the sources per role and an audit log that matches the access-logging standard, exportable to your own monitoring. See AI for healthcare.
Frequently asked questions
Questions on this topic
Does NEN 7510 say anything about AI?
Not literally. But every AI application that processes patient data falls under the management system, like any other system: risk assessment, access, logging, supplier management and incident handling.
What is new in NEN 7510:2024?
The December 2024 version is based on ISO 27001:2022 and ISO 27002:2022, prescribes less and asks for more of the organisation’s own risk assessment. Certified organisations must have switched before 20 February 2027.
Is a cloud AI service allowed in healthcare?
That depends on the processing agreement, the location of the data, the exclusion of training and the logging. A node that belongs to the institution alone makes those questions easier to answer than a shared service.
How does AI logging match the access-logging standard?
NEN 7513 asks that access to the patient record be traceable. An AI assistant that reads records must therefore record per question who, when and which record, and that log must be kept and searchable.
Sources
- NEN — Information security in healthcare (NEN 7510, Dutch)
- Dutch Health and Youth Care Inspectorate (IGJ)
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.