Skip to content

sandbox.md

Latest commit

 

History

History
101 lines (74 loc) · 2.77 KB

sandbox.md

File metadata and controls

101 lines (74 loc) · 2.77 KB
Raw

Sandboxing

The agent has a shell. On a machine that also holds your files, your keys, and your network, that is a lot of reach for something driven by a model following instructions from a web page it just read.

terminal:
  sandbox: auto        # none | auto | bubblewrap | namespace
  allow_network: false

Nothing here needs root.

What each mode does

Mode Filesystem Network Needs
none everything you can reach open —
namespace unchanged gone Linux user namespaces
bubblewrap read-only except the workspace, credentials hidden gone unless allowed bwrap installed
auto the strongest of the above that works here —

auto is the setting to use. It picks bubblewrap when installed, falls back to namespaces, and says which it got — once, in the log, not on every command.

Namespaces

Available on any Linux kernel with unprivileged user namespaces enabled, which is most of them. The command runs in its own user, mount, PID, IPC, and UTS namespaces, and — unless allow_network is true — its own empty network namespace.

There is no filesystem confinement this way. Taking the network away is still most of the value: an instruction the agent picked up from a page it read cannot send anything anywhere.

Bubblewrap

apt install bubblewrap     # or the equivalent

Then the filesystem is confined too:

  • everything readable, nothing writable except the workspace
  • a private /tmp
  • credential directories replaced with empty ones
terminal:
  sandbox: bubblewrap
  sandbox_hidden:
    - ~/.ssh
    - ~/.aws
    - ~/.gnupg
    - ~/.config/gh
    - ~/.antares/.env
    - ~/.kube
    - ~/.docker/config.json

That list is the default. Setting sandbox_hidden replaces it rather than adding to it.

Docker

For stronger separation, run commands in a container instead:

terminal:
  backend: docker
  docker_image: debian:bookworm-slim
  allow_network: false

The workspace is mounted; nothing else is. This is the strongest option, and the slowest to start.

When it cannot be built

A missing mechanism downgrades and logs why, rather than refusing to run anything — a sandbox you cannot have should not stop you working. Which means you should check what you actually got:

antares doctor

A sandbox you believe in but do not have is worse than none, so the mode is reported rather than assumed.

What it does not do

It does not stop the agent from doing damage inside the workspace — that is what approvals and checkpoints are for.

It does not filter the network, only remove it. There is no allow-list of hosts.

It does not confine the file tools, only the shell. Those are already bounded by the workspace.