Sandbox

Every other tool is bounded by the path jail: ToolCtx::resolve proves a path stays inside the workspace before anything touches it. shell cannot work that way. The jail cannot see inside bash -c, and a command is free to cd /, read ~/.aws/credentials, and curl it out.

The capability model has always said as much — shell declares private_data, destructive, and chosen egress — but saying it is not enforcing it. sandbox.rs is the enforcement.

Configuration

[sandbox]
kind = "bwrap"              # none | bwrap | docker | landlock
network = false             # confined commands reach the network
writable = []               # extra paths mounted read-write
readable = ["/opt/toolchain"]   # extra paths mounted read-only
env = ["CARGO_HOME"]        # environment variables passed through, by name
image = "debian:stable-slim"    # docker only
memory_mb = 2048            # docker only
cpus = 2.0                  # docker only

kind	What it does
none	The default. Commands run directly, as you, with your credentials. The only sane default for a supervised CLI on a machine where the alternatives may not be installed.
bwrap	User namespaces via bwrap. Cheap — no daemon, a few milliseconds per command.
docker	A throwaway container (--rm). Works where user namespaces are locked down; costs a container start per command.
landlock	Landlock LSM rules applied in the child between fork and exec. Needs no privilege at all, so it works where bwrap fails even installed (Ubuntu 23.10+'s AppArmor userns switch). Kernel 6.2+ required — older ABIs cannot restrict truncation, and half-confinement is refused rather than served.

mecha tools prints the active sandbox, including whether the network is on and whether reads can leave the workspace.

What landlock deliberately does not claim

Landlock confines files: workspace writable, system read-only, your home directory denied wholesale. It cannot close the network — TCP is denied on 6.7+ kernels as defense in depth, but UDP is unrestrictable at any ABI, and echo x > /dev/udp/host/port works in bash alone. So a landlocked shell never earns the trifecta interlock's relaxation: its egress stays chosen whatever network is set to, and the denial message says so rather than pointing at a setting that would not help. Its preflight proves the denial, not just the apply: it plants a file in your real home and requires the confined read to fail — "confined" with nothing denied is the state the preflight exists to forbid. Three known ways it is weaker than bwrap: /tmp is shared rather than private, /proc shows every process, and there is no PID/IPC isolation.

What a confined command gets

• The workspace, mounted writable, plus anything named in writable.
• A read-only system — /usr, /etc, /opt, and the /bin, /lib family (bound as directories or recreated as symlinks, depending on whether the host has a merged /usr), plus anything named in readable.
• A private /tmp, so one command cannot leave anything behind for the next or read what the last one left.
• No home directory. HOME is set to the workspace.
• No environment except an allowlist. bwrap gets --clearenv, then a fixed PATH, then only the variables named in env. API keys live in the environment; a confined command that inherits them is confined in the least interesting way.
• No network, unless network = true.

Under bwrap the process also gets --die-with-parent (a wedged command cannot outlive mecha) and --new-session (which detaches the controlling terminal, blocking the TIOCSTI trick of pushing characters into the parent's input queue). Under docker it runs as your uid/gid — root inside the container writing into a bind-mounted workspace would leave root-owned files you then cannot delete — with --cap-drop ALL and --security-opt no-new-privileges.

The same machinery confines MCP servers, which need it more: an MCP server is third-party code running on your machine, where shell at least runs commands a model asked for out loud. wrap_argv confines an explicit argv with no shell in between, because quoting caller-supplied arguments into bash -lc is a command-injection bug waiting for one entry with a space. A server's network setting overrides the global one, so one server can reach its own API without opening shell to satisfy it.

This is not a boundary against a determined kernel exploit. It is the difference between "an injected command can read your SSH keys" and "an injected command can read the files you pointed the agent at" — which is the difference that decides whether an agent can safely be woken by an email.

Three rules

1. A configured sandbox that does not work stops the run

Sandbox::preflight runs a real command through the real backend at startup and checks for its marker in the output. A failure is an error with instructions, not a warning, and shell is refused rather than run unconfined.

sandbox preflight failed — refusing to run `shell` unconfined

Silently falling back to unconfined execution is worse than having no sandbox at all, and not only because the operator writes policy on a false belief: shell declares narrower capabilities when confined, and the interlock believes it. A degraded sandbox would leave the loop treating a fully capable shell as one that cannot reach the network.

The check happens once at startup rather than on the first tool call, so a misconfiguration is a clear message at launch instead of a confusing tool error twenty turns in. If the sandbox cannot be built at call time either, the call is refused with "Nothing was executed" rather than falling through.

2. Only egress narrows; private_data stays true

Capabilities {
    private_data: true,
    untrusted_input: false,
    // Confinement removes the route rather than narrowing its class: an
    // unconfined shell can `curl` any host the model names, which is the
    // complete channel.
    egress: if self.sandbox.can_reach_network() { Egress::Chosen } else { Egress::None },
    destructive: true,
}

No network means no way out, so a confined shell genuinely stops being a trifecta sink and the interlock stops refusing outbound-looking work that provably cannot go anywhere. That relaxation is earned by preflight: the label narrows only because something else enforces it.

private_data does not narrow. A confined shell still reads the workspace, and fs_read — which reads exactly the same bytes — is marked private on the grounds that your files are the definition of private data. Narrowing it here would open a hole rather than close one: shell: cat secrets would set no taint where fs_read: secrets does, and the cheapest route around the interlock would be to use the more dangerous tool.

Note that an extra readable or writable bind puts data back in reach. reaches_beyond_workspace() reports true whenever either list is non-empty, because an extra bind is exactly how private data gets back in scope.

3. The policy lives on the tool, not in ToolCtx

capabilities() takes no context, so it has nothing to consult — the sandbox is handed to the tool when the registry is built:

let registry = Registry::new().with_builtins(&cfg.tools, Arc::clone(&sandbox));

The workspace still comes from ToolCtx at call time, so a per-run jail — an eval case's private fixture copy, a batch item's directory — is what actually gets mounted. Confinement policy is fixed per process; what it is pointed at is per run.

Ubuntu 23.10 and later

warning

On Ubuntu 23.10+, bwrap fails even when it is installed and kernel.unprivileged_userns_clone is 1, because AppArmor gained a separate switch that nothing mentions:

sysctl kernel.apparmor_restrict_unprivileged_userns   # 1 means blocked

Either install an AppArmor profile for bwrap, set kernel.apparmor_restrict_unprivileged_userns=0 (system-wide, and weaker), or use kind = "docker" instead.

preflight recognises this case and prints that advice with the failure, along with the other common ones: a bwrap that cannot configure loopback in a new network namespace, and a docker socket you do not have permission to use.

Where to go next

• Security model — what the capability labels feed.
• Tools and MCP — confining servers, and the environment allowlist.
• Configuration reference — every [sandbox] key.