When agents run untrusted code at scale, containers that share a host kernel may not be enough. gVisor intercepts system calls in a user-space kernel, and Firecracker runs lightweight virtual machines. Both trade some compatibility and speed for a much smaller attack surface.
Read in this order
- ArticlePractical security guidance for sandboxing agentic workflows (NVIDIA AI Red Team)Start here: when container isolation falls short for agent workloads.
- OfficialgVisor documentationHow a user-space kernel narrows the host attack surface.
- VideoBeyond Containers: MicroVMs as the Security Primitive for AI Agents (The Context)Demonstrates Micro Sandbox and its hardware isolation boundary. 25 min.