Mastercard says AI agents need 'confidential computing' because their memory is the new target

As AI agents become a standard part of enterprise operations, especially in sensitive sectors like finance, a new security frontier is opening up. Mastercard's senior vice president of cloud engineering, Krishna Mayuram, explains that while data at rest and in transit are already protected, the data actively being processed by AI agents—their 'memory'—now represents the primary vulnerability. These agents, running on Kubernetes platforms like Red Hat OpenShift, handle customer transaction data under strict regulatory oversight across multiple countries.

Mayuram argues that traditional encryption isn't enough; the industry needs 'confidential computing,' which involves running AI workloads within highly secure, virtualized environments or confidential containers. This approach protects data while it's actually being used, not just stored or transferred. For engineering teams at companies like Mastercard, scaling these AI agent platforms securely requires building guardrails at every stage, from development to deployment, to meet diverse legal requirements and protect against the heightened risks that come with distributed, agentic systems.