Shared kernels start lighter; MicroVMs add a hardware boundary
A Sandbox uses namespace, seccomp, subordinate IDs, and cgroup v2 process isolation, so it does not boot a guest kernel. The agent still shares the host Linux kernel: if a working host-kernel exploit succeeds, process isolation may fail with it. The default MicroVM boots a dedicated guest kernel, so taking over that guest kernel still leaves a hardware VM boundary to cross.
Creating a MicroVM has higher startup and memory cost because it starts a VMM, guest kernel, and guest init. Warm pools and snapshot-fork can pay that cost before a task arrives. The hypervisor, hardware, mounts, network exposure, and side channels still require review.a3s-box run --rm alpine:3.20 -- uname -a