OS Sandbox
Kernel-enforced containment for agent shell commands — Seatbelt on macOS, bwrap on Linux
The OS sandbox puts a real /bin/sh -c inside a kernel-enforced filesystem jail — sandbox-exec (Seatbelt) on macOS, bwrap (bubblewrap) on Linux — and refuses to run unless it has measured that the jail actually blocks writes. Under PRAISON_SHELL=sandboxed this applies to every execute_command, not only commands with shell operators.
Quick Start
- Run the coding agent sandboxed
PRAISON_SHELL=sandboxeddrives the OS sandbox — a real shell inside Seatbelt or bwrap.bash PRAISON_SHELL=sandboxed praisonai codetext >>> run: echo written > r2.txt exit=0 success=True sandbox=seatbelt; r2.txt contents 'written' - Watch an escape get denied by the kernel
A write outside the workspace hits a real OS denial, not a soft warning.
text >>> run: echo pwned > /tmp/PWNED.txt /bin/sh: /tmp/PWNED.txt: Operation not permitted PWNED exists: False
Two Backends
Each platform uses its native primitive; everywhere else, no wrapper is built.
| Platform | Backend | How it confines |
|---|---|---|
| macOS | sandbox-exec (Seatbelt) | Generated profile denies file-write* outside the writable set, and denies reads and writes under the masked secret dirs |
| Linux | bwrap (bubblewrap) | Read-only root, workspace bind-mounted writable, --tmpfs over the masked secret dirs, --unshare-pid, network unshared |
| Other | none | build_wrapper() returns None — containment unavailable |
Workspace and cwd
The writable set is PRAISONAI_WORKSPACE (or the process cwd) plus $TMPDIR. The tool’s cwd= argument must be inside that set, or the command is refused — the model cannot promote a path to writable by passing it as cwd. Network is denied by default.
Environment and masked paths
Env scrub. Variables whose names match KEY|TOKEN|SECRET|PASSW|CREDENTIAL (case-insensitive) are not passed into the jail. Entries passed explicitly through execute_command(env=...) are the allow-list and still go through.
Env applied inside the jail. Explicit env= entries are set by /usr/bin/env k=v … inside the jail, never on the bwrap / sandbox-exec launcher — otherwise LD_PRELOAD and friends would load code before containment starts.
Masked dirs. ~/.ssh, ~/.aws, ~/.gnupg, ~/.config/gcloud and ~/.azure are hidden (--tmpfs on bwrap, deny file-read file-write on Seatbelt). The rest of the filesystem stays readable because toolchains need it.
Own PID namespace (Linux). --unshare-pid makes the fresh /proc show only the jail, so /proc/<host pid>/environ cannot reveal the scrubbed env.
Refuse, Never Degrade
A sandbox that is assumed but not enforcing is worse than none — callers relax on a belief that isn’t true.
PRAISON_SHELL=sandboxed refuses to run when the OS primitive is missing or not enforcing, rather than silently falling back to an uncontained shell. To accept an uncontained shell deliberately, use PRAISON_SHELL=unsafe.
Probe Before Claiming
probe_enforcement() measures the jail instead of trusting it.
The probe runs a child under the wrapper that writes an in-jail canary and attempts a write outside the writable set:
- If the escape write succeeds, the backend is reported unavailable.
- The canary guards against a false positive — if the wrapper couldn’t start the child at all, the missing escape file proves nothing, so the canary must exist for the result to count.
Results are (enforcing: bool, backend: str, detail: str) and cached per process. Flipping to a permissive Seatbelt profile flips the probe from (True, 'seatbelt', 'blocked …') to (False, 'seatbelt', 'did not block a write outside the writable set').
Not SandboxConfig.native()
Best Practices
Trust the refusal
If sandboxed mode refuses because containment can’t be proven, install the primitive (bwrap on Linux) rather than switching to unsafe.
Scope the writable set
Set PRAISONAI_WORKSPACE to the project root; only it and the temp dir are writable. Keep other paths out so an escape attempt is a real out-of-jail write the kernel can deny.
Leave network off unless needed
network=False unshares the network on Linux and denies it in the Seatbelt profile. Enable it only when a build step genuinely needs it.
Pair with approval
The sandbox runs behind the critical-risk execute_command approval — containment and human sign-off reinforce each other.