Case study · Healthcare security

One front door for research: securing lab-deployed AI tools at a Japanese university hospital

A Japanese university hospital ran research AI tools that individual labs had installed themselves: imaging annotation servers, notebook servers, model demos. Some were reachable from the internet, outside the hospital's security operations. When AI agents were reported to have found and exploited a zero-day in a medical imaging tool on exactly that kind of server, the goal became clear: find every research service, put it behind one front door, and do it without stopping the research.

Case studyHealthcare and research · JapanSumanta Mukhopadhyay9 min read
#healthcare#security#kubernetes
CASE STUDY · CASE STUDY · HEALTHCARE SECURITY22 Nov 2026 · 15:50 GST
At a glance
SectorHealthcare and research · Japan
My roleFractional CTO and security architect
EngagementRapid assessment, architecture and delivery leadership
ScopeResearch service discovery, central gateway, identity, network segmentation, research data handling, monitoring
What changed
  • Every internet-reachable research service found and assigned an owner
  • Research tools published only through one authenticated gateway
  • Labs able to deploy their own tools within platform guard rails
  • Patient data kept in a separate zone from experimental tools
  • Security monitoring covering research services for the first time
External attack surface scanningKubernetes research platformGateway APIIdentity-aware accessHospital identity providerNetwork segmentationDe-identified research data zoneSecurity monitoring

Anonymised composite. Client names, identifying details and confidential figures are withheld; public facts are cited in the references.

The situation

A Japanese university hospital had a strong central IT function for clinical systems and a very different world in its research labs. Radiology, pathology and informatics groups ran their own AI tools: imaging annotation servers for labelling studies, notebook servers for model development, demo applications for collaborators. Some sat on lab machines, some on cloud accounts paid from research grants, and some were reachable from the internet so partners at other institutions could log in.

The hospital followed the Ministry of Health, Labour and Welfare's guidelines on the safety management of medical information systems, whose version 6.0 was restructured to respond to cyberattacks and wider cloud use, with separate volumes for governance, management and system operations3. Those controls were well applied to clinical systems. Research services were largely outside them.

Why now. In November 2026, Amazon reported that a suspected China-nexus operator used autonomous AI agents to map the API of MONAI Label, an open-source medical imaging annotation tool, find a path traversal flaw and exploit it on university servers in East Asia1. The Register reported the agents went on to gain root2. The hospital's research groups used tools of exactly this kind. The board asked how many such servers were exposed. Nobody could answer.

What had to be true

  • Every research service known, with an owner, purpose and the data it can reach.
  • Nothing research-related directly reachable from the internet; one authenticated front door instead.
  • External collaborators still able to work, with their own identities, not shared passwords.
  • Patient data separated from experimental tools, with only de-identified data in research zones by default.
  • Research not stopped: labs keep the freedom to deploy, inside guard rails.

Architecture

Users
Hospital researchershospital identityExternal collaboratorsfederated or guest identities
Front door
Identity-aware gatewaysingle public entry pointPer-route access policywhich group sees which tool
Research platform
Kubernetes namespaces per lablabs own their routesAnnotation and notebook servicesapproved imagesGPU poolshared, scheduled
Data
De-identified research zonedefaultIdentifiable dataseparate, approval requiredImaging archiveclinical, read-only exports
Assurance
External scanningcontinuousSecurity monitoringgateway and platform logsService registerowners and reviews
One public entry point, labs owning their own routes behind it, and data zones that keep experiments away from identifiable records.

The data flow, end to end

  1. DiscoverExternal scans and cloud account reviews find research services
  2. RegisterEach one gets an owner, a purpose and a data classification
  3. MoveServices move onto the research platform or are retired
  4. PublishAccess only through the gateway, with authentication per route
  5. FeedData reaches tools from the de-identified zone by default
  6. WatchGateway and platform logs go to the hospital's monitoring
Find, register, move, publish safely, then watch: the same path for every research tool.

The hard part: central control without central bottlenecks

The easy answer, making central IT deploy every research tool, would have stopped research and pushed labs back to their own servers. The design had to separate who controls the front door from who controls each service. Kubernetes Gateway API is built around exactly that split: infrastructure providers define gateway classes, cluster operators run gateways, and application developers attach their own routes5.

Diagram of Gateway API roles: an infrastructure provider manages the GatewayClass, a cluster operator manages the Gateway, and application developers manage HTTPRoutes that point to Services
Role-oriented routing: the platform team owns the gateway, each lab owns its routes. Source: Kubernetes blog, Lior Lieberman and Kobi Levi · CC BY 4.0

In practice, the hospital's platform team ran the gateway and its authentication; each lab had a namespace and could publish routes to its own tools, which the gateway only accepted with an access policy attached. A new annotation server took the same steps for a lab as before, but could no longer appear on the internet without login.

The fastest win came before any platform work: an external scan of the hospital's address ranges and research cloud accounts. It turned an unknown into a list with owners.

Data was the second half. The ministry also publishes cyber security measures for medical institutions alongside the guidelines4, and the hospital extended the same thinking to research. Research tools received de-identified exports by default; identifiable data stayed in a separate zone that required approval and was never reachable from tools exposed to external collaborators. The pattern follows the health evidence platform and federated health analytics, and the gateway approach builds on moving to Gateway API.

How it ran

Information security
  1. External scans
  2. Service register and reviews
  3. Monitoring
Research platform team
  1. Gateway and identity
  2. Lab namespaces and approved images
  3. GPU scheduling
Research labs
  1. Confirm owners and purpose
  2. Move or retire services
  3. Publish routes with access policies
Security found the services, the platform team built the front door, labs moved their own tools.
Phase 1Find
  • External and cloud scans
  • Owners for every service
  • Close the worst exposures
Phase 2Front door
  • Identity-aware gateway
  • Collaborator identities
  • Route policies
Phase 3Platform
  • Lab namespaces
  • Approved images
  • Data zones
Phase 4Sustain
  • Continuous scanning
  • Monitoring
  • Periodic reviews
Close exposures first, then give labs a better way to deploy than the one they had.

Key decisions and trade-offs

DecisionAlternativeWhy
One authenticated gateway for all research toolsSecure each tool individuallyTools vary too much; the front door is the control point
Labs own their routesCentral IT deploys everythingKeeps research moving; avoids new shadow servers
Federated identities for collaboratorsShared accounts and passwordsAccess tied to a person and revocable
De-identified data by defaultIdentifiable data where convenientLimits harm if a research tool is compromised
Retire what has no ownerLeave running until someone complainsUnowned services are never patched

What I'd do again, and differently

Again: start with an external scan and a service register; it changed the conversation from fear to a list. Differently: bring lab leads in as co-designers of the platform from the first week. Researchers adopted it fastest where they had shaped the deployment templates. See AI-assisted intrusions and the tech news for 22 November.

If you run healthcare or research technology in Japan or across Asia, see my regional work, or analyse your challenge and bring the result to a call.

References
  1. 1
  2. 2
  3. 3
  4. 4
    Cyber security measuresMinistry of Health, Labour and Welfare (Japan)
  5. 5
    Introducing ingress2gateway: Simplifying Upgrades to Gateway APIKubernetes blog (Lior Lieberman, Kobi Levi) · 25 Oct 2023
FAQ

Common questions

What happened with MONAI Label?

Amazon reported that a suspected China-nexus operator's autonomous AI agents mapped the API of MONAI Label, an open-source medical imaging annotation tool, found a path traversal flaw and exploited it against university servers in East Asia.

Why are research tools a particular risk in hospitals?

They are often installed by individual labs, outside central IT, on servers that are reachable for collaborators, and they can sit near imaging or clinical data. They are rarely patched or monitored like clinical systems.

Does this mean banning lab-deployed tools?

No. The aim is a platform where labs can still deploy what they need, but only behind central authentication, on a segmented network, with monitoring and an owner for every service.

Keep reading

Related

Open to advisory, fractional CTO & select builds

Let's talk about what you're building.

The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.

Start a conversation

How would you like to reach me?

Book a 30-min callFastest. Pick any open slot on my calendar.→ Write in GmailOpens a pre-filled draft in your browser.→ Use my email appOutlook, Apple Mail or your default client.→

I reply personally within two business days. Time zones from the Americas and Europe to the Middle East and Asia-Pacific.