3-minute product tour Urgent incident email
Security blog

Tenant data isolation in security AI

Security AI processes logs, vulnerability data, and internal network configuration.
Public research on tenant isolation informs the operational and data-handling requirements for dedicated customer environments.

Tenant data isolation in security AI

Data handled by security AI

Security AI brings together information that intruders would otherwise have to gather over weeks. It can reveal which servers remain vulnerable, where administrators sign in, and how the internal network is segmented.

Tenant isolation failures in AI platforms

Wiz Research reported tenant isolation failures on two AI service platforms in 2024. It found that Hugging Face inference infrastructure could expose other customers’ models after a container boundary failed, and that Replicate could expose other customers’ prompts and results. Both companies worked with Wiz to fix the issues.

Both cases involved customer-supplied models or containers running on shared infrastructure. This is common in AI services. If multiple customers’ workloads share a cluster, data safety depends on how effectively it enforces tenant boundaries.

How Genox separates customer environments

We separate compute and networks for each customer. Subsystems run on designated nodes reserved for their workloads, and network policy restricts outbound traffic. A Kubernetes deployment can use the structure shown below.

kubectl taint nodes gx-node-01 dedicated=tenant-a:NoSchedule

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: default-deny-egress
  namespace: tenant-a
spec:
  podSelector: {}
  policyTypes: ["Egress"]

Taints keep unrelated pods off dedicated nodes. Use a matching toleration and nodeAffinity so the intended pods are scheduled there. Otherwise, the dedicated nodes can remain empty while those pods run on shared nodes. Verify NetworkPolicy enforcement after applying it, since it depends on support from the cluster’s network plugin.

We default to isolated on-premises deployments. For cloud deployments, we provision a dedicated customer environment in a Korean region.

The operational costs

Dedicated environments cost more because resources remain reserved even when idle. Isolated networks cannot call externally hosted models, and model and detection-rule updates must pass through controlled import and export procedures, which can slow rollouts.

We consider these costs justified given the data involved. The right choice depends on the organization’s data, regulatory requirements, and available operations staff.

References

Talk to Genox
about scope and timing.

Talk to us