Isolation
A virtual machine per workspace
Section titled “A virtual machine per workspace”Every workspace is a QEMU/KVM virtual machine with a kernel of its own. Workspaces share no kernel with each other or with the runner. Code that takes root in a workspace, or its kernel, is still inside a virtual machine.
The QEMU Spinbox ships is built for KVM only: it refuses to emulate.
QEMU itself is confined
Section titled “QEMU itself is confined”If a workspace escapes into its QEMU process, that process reaches almost nothing:
- It runs as a user of its own, not as the runner’s user.
- It starts in a network namespace with nothing in it. It cannot open a connection to the runner, another workspace or the network.
- Its vsock is in
localmode, so it cannot reach another machine’s vsock. The guest answers only on a reserved vsock port, which only the runner binds. A machine’s vsock number names its workspace only while that machine holds it, and a terminal or a port is opened only while the number still names the same workspace, so a number handed to the next machine does not carry the last one’s sessions with it. - It is born in a cgroup with
memory.maxandpids.max, and refused if one cannot be made. - It has a core-scheduling cookie of its own for every thread, so its vCPUs never share a physical core’s siblings with another machine’s, where one could read what the other leaves in the core’s caches.
- Same-page merging is off on the runner, so one machine cannot learn another’s memory by timing page merges.
- Kernel modules for socket families nothing uses are never loaded.
- It gets no environment, and every message it sends on its control socket is bounded before it is decoded.
Untrusted bytes are parsed confined
Section titled “Untrusted bytes are parsed confined”A workspace’s disk is bytes its tenant wrote as root. The runner never mounts or parses a guest’s filesystem:
qemu-img, the one program on the runner that reads a workspace’s disk images, runs with no capabilities, under Landlock limited to the files of that one operation, under seccomp with no sockets, in namespaces of its own.- A diff between two snapshots runs in a separate virtual machine with no network, whose kernel is the only one that parses the tenant’s filesystem. Its answer is checked and bounded before it is believed.
- Text a guest sends to the host — console output, names, paths — is bounded in size, count and rate, and made valid, where it crosses.
The network boundary
Section titled “The network boundary”The workspace’s network is enforced in nftables on the runner, in a namespace of its own, outside the virtual machine. It drops every packet to private, loopback, link-local (including the instance metadata service) and reserved addresses, and to the runner’s own addresses, before any policy is read. The workspace’s resolver removes those addresses from its answers.
A workspace is never told the control plane’s address or its CA. The programs Spinbox runs inside it reach the control plane over vsock, through the runner.
Where it depends on the kernel
Section titled “Where it depends on the kernel”Three of these protections need the runner’s kernel to support them: vsock namespaces, core scheduling and Landlock. A runner says what its kernel keeps when it registers, and a production installation places no workspace on one that lacks any; only a development installation starts machines anyway. See Known gaps.
What this is, and isn’t
Section titled “What this is, and isn’t”Each line above is a mechanism with code and a test behind it, not a promise that nothing gets through. A boundary is only as good as the last serious attempt to cross it, and the models that run in workspaces keep getting better at finding doors. The open items are on Known gaps.
Installations run Spinbox v20261027.01; the reference is generated from spin v20261114.01 (51f7e0dea). This page as Markdown, for a model or a terminal.

