Video•Feb 24, 2026
AI Agent Sandboxes: Securing Memory, GPUs, and Model Access
AI agents expand the attack surface beyond traditional microservices by accessing models, memory, GPUs, and external tools, making conventional container isolation inadequate. Industry leaders discussed how lightweight VM‑based sandboxes, such as those built on Kata, can provide stronger isolation, limit host visibility, and mitigate risks like GPU memory leakage. The session covered runtime guardrails, privilege boundaries, telemetry, and performance trade‑offs, framing secure‑by‑design as essential for production AI workloads. Attendees debated whether agent virtualization will become a standard layer in AI infrastructure.
