ABOUT THIS EPISODE
AI agents introduce a fundamentally different risk profile than traditional microservices or batch workloads. With access to models, memory, tools, external APIs, and sometimes direct execution capabilities, agents can observe, reason, and act in ways that expand the attack surface far beyond standard containerized applications. The question is no longer just how to scale AI, but how to securely contain it.In this live discussion with leaders from Ant Group, NVIDIA, Google, and the Linux Foundation, we’ll explore why traditional container isolation may be insufficient for agent-based systems, and what changes when agents have memory persistence, filesystem access, GPU acceleration, or external execution authority. We’ll examine how approaches like Kata-based agent sandboxes provide lightweight VM isolation to restrict runtime behavior, minimize host visibility, and reduce cross-session risk, including emerging concerns like GPU memory leakage.From runtime guardrails and privilege boundaries to telemetry capture and performance trade-offs, this session will unpack what “secure-by-design” means for AI agents. Finally, we’ll look ahead: will agent virtualization become a standard layer of AI infrastructure, and are sandboxes destined to become a universal requirement for production AI platforms?
