Skip to the yard
Waxsealai

Confidential Computing Targets AI Agent Data on Kubernetes

Shipped
Shipper
Jonathan Anthony
Readers
714

Mastercard’s cloud engineering approach highlights why agent platforms need protection for data in use alongside Kubernetes scale and enterprise guardrails.

Confidential Computing Targets AI Agent Data on Kubernetes

AI agents running on Kubernetes create a security gap that encryption for stored and transmitted data does not address: information can still be exposed while a workload is processing it. For companies operating transaction platforms across many jurisdictions, that gap makes confidential computing part of the infrastructure discussion, alongside the ability to scale agent workloads.

Krishna Mayuram, Mastercard’s senior principal engineer and vice president of cloud engineering, describes confidential computing as a way to protect data during processing through secure virtualized environments or confidential containers. The distinction is practical. Encryption can protect information at rest and in transit, but applications must often work with usable data while they run. Mayuram’s question is what protections apply at that point.

Protecting data while agents run

Agents behave much like microservices in the infrastructure model Mayuram outlined. They need elastic capacity, and Kubernetes can provide horizontal and vertical scaling for both the platform and the workloads it hosts. That lets teams adapt computing resources as demand changes, but expanding the number of running workloads also expands the environment that must be managed and protected.

Confidential computing addresses one part of that problem by aiming to isolate processing environments and protect the data handled inside them. It does not remove the need for sound application design or operational controls. Rather, it adds a layer directed at exposure during execution, when software needs access to information to perform its tasks.

Mastercard runs its platforms on Red Hat OpenShift, a Kubernetes platform. Red Hat has extended OpenShift to support AI agents and has introduced confidential containers for bare-metal deployments. Those capabilities put confidential workloads within a Kubernetes operating model, where teams can manage agent services using familiar orchestration patterns.

For infrastructure teams, that combination matters because security and scaling have to work together. A protection mechanism that cannot fit the way workloads are deployed and expanded may be difficult to apply consistently. Conversely, scaling agents without accounting for what data they process can multiply exposure as workloads spread across a platform.

Guardrails across deployment

Mayuram said Mastercard’s approach applies guardrails from development through deployment. He argued that the basic enterprise security requirements remain the same whether a workload uses agents or conventional software. The distributed nature of the business and the variety of rules that apply across regions make those controls necessary at each stage.

That framing puts agent security within existing engineering practice rather than treating it as a separate category. Teams still need to define which data a workload can access, how it is deployed, and what constraints apply as it runs. Confidential computing can strengthen the execution environment, while platform controls and development practices govern the broader lifecycle.

Mayuram also connected this work to Mastercard’s services for agentic commerce. The company recently added trust services for that area; internally, he said, the same emphasis on guardrails applies. The comments point to a common operational challenge: as software takes on more autonomous tasks, organizations need clear boundaries around the data and infrastructure available to it.

What platform teams need to assess

The discussion does not provide measurements of confidential computing’s performance overhead, deployment cost, or effect on agent throughput. Those details matter for teams deciding where to use protected execution environments. They will need to evaluate how the chosen platform supports their workloads, what operational changes are required, and whether the protection fits the sensitivity and processing needs of the data involved.

Nor does confidential computing replace the controls that govern an agent’s behavior or access. An isolated environment can help protect information during processing, but organizations still need to manage permissions, deployment changes, and the boundaries between workloads. The source of the security benefit is narrower and more concrete: reducing exposure of data while software uses it.

The practical takeaway is that Kubernetes agent adoption brings two linked infrastructure requirements: elastic capacity and controls that cover data in use. For teams building these systems, the next step is to assess how confidential environments fit their existing platform and lifecycle guardrails, then determine which workloads and data warrant that added layer.

Source: Jonathan Anthony · siliconangle.com

Same vesselRelated cargo