grok-build-upstream-mirror/crates/codegen/xai-grok-pager/docs/user-guide/18-sandbox.md
grokkybara[bot] 69f0ba880a Synced from monorepo
Synced from monorepo

Changes:
- Workspace server: report `/ready` as failed with dwell on hub connect failure
- Refresh OIDC token for the Grok agent in the shell
- ACP terminal output recorder
- Cross-platform provider auth commands in the shell
- Default `/resume` to Grok sessions with a hint for hidden external sessions
- Resume sessions by title with `--resume`
- Limit app-builder archive size
- Data-driven tag labels for slash commands
- Doctor fixes for tmux
- Custom provider gateways and subprocess environment policy in the shell
- `/tutorial` — opt-in onboarding tour of Grok Build
- Soft and required CLI version checks in the shell
- Privacy banner env overrides survive live settings updates
- Add remote flag to override the image-edit model
- Return profile fields from auth info even when the access token is expired
- Add edit control on queued prompt rows
- Keep fail-closed policy when clearing orphans with no team
- Setting to disable the Ctrl+Space/F8 voice shortcut
- Pass `--raw` to pw-record so Linux dictation works on older PipeWire
- Validate git URLs when adding marketplace entries
- Stop shipping stale tool-doc parameter and tool names
- Re-point dashboard attach after `/fork` only when the parent was attached
- Surface Grok Computer media-generation results as file-path chunks
- Clear web background-task tray on kill and keep the task description
- Show privacy upsell banner in agent view until acted on
- Add tools-server client callback surface
- Protect persistent global hook sources

Source-Revision: 95d84f443eddcbed6cbfd6eed22e2eafe6b3939d
2026-07-23 17:12:33 +00:00

16 KiB

Sandbox Mode

Sandbox mode restricts what the agent process and its spawned commands can access on your filesystem and network using OS-level kernel primitives (Landlock on Linux, Seatbelt on macOS). The kernel enforces these limits for the process lifetime.

Sandbox mode is off by default.


Quick Start

# Run with workspace sandbox (read everywhere, write to CWD + temp dirs + ~/.grok/)
grok --sandbox workspace

# Read-only mode (read everywhere, write only to ~/.grok/ + temp dirs)
grok --sandbox read-only

# Most restrictive profile (read CWD + system paths, write CWD + temp dirs + ~/.grok/, no child network)
grok --sandbox strict

Built-in Profiles

Profile FS Read FS Write Child Network Use Case
off (default) Unrestricted Unrestricted Unrestricted No sandbox
workspace Everywhere CWD + ~/.grok/ + /tmp + /var/tmp Allowed Normal development
devbox Everywhere All top-level dirs except /data Allowed Disposable dev VMs
read-only Everywhere ~/.grok/ + /tmp + /var/tmp Blocked¹ Exploration, code review
strict CWD + system paths CWD + ~/.grok/ + /tmp + /var/tmp Blocked¹ Untrusted code

¹ Child-network blocking is enforced on Linux only (via seccomp). On macOS it is a no-op — these profiles do not restrict child-process network there.

To block specific files (e.g. .env or credential paths) on top of a profile, define a custom profile with a deny list — it is kernel-enforced (read + write/rename) and supports glob patterns like **/*.pem.

Profile Details

workspace -- The recommended profile for everyday development. The agent can read any file on the system (for understanding dependencies, system libraries, etc.) but can only write to the current working directory, ~/.grok/, and temp directories (/tmp, /var/tmp, plus the macOS temp dirs). Network access is allowed for tools like web_search and MCP servers.

devbox -- A reserved built-in profile for disposable development VMs. The agent can read everywhere and write to every top-level directory except /data and the virtual filesystems (/proc, /sys, /dev), including the home directory. Network access is allowed. --sandbox devbox runs the built-in profile, which shadows any [profiles.devbox] you define in sandbox.toml.

read-only -- Use when you want the agent to analyze code without modifying your project files. The agent can read everything but can only write to ~/.grok/ (needed for session persistence) and temp directories. Child-process network access is blocked on Linux (no-op on macOS).

strict -- The most restrictive profile, for reviewing untrusted code. The agent can only read files within the current working directory and essential system paths. Writes are limited to CWD, ~/.grok/, and temp directories. Child-process network access is blocked on Linux (no-op on macOS).

Direct global hook write protection

Under workspace, read-only, and strict (and custom profiles that extend those bases), the Grok state directory remains writable for session/runtime files, but the kernel write-denies the Grok-owned direct disk paths used as user-global hook sources (they stay readable):

  • ~/.grok/hooks/ (hook directory)
  • ~/.grok/hooks-paths (registry file; not loaded as hook JSON — only its absolute targets are)
  • Absolute targets listed in hooks-paths (relative lines are ignored; missing targets refuse sandbox start)

On first launch under these profiles, Grok creates a real empty hooks/ directory and empty hooks-paths file when they are missing (never symlinks or wrong types). Claude/Cursor global settings are not covered by this write-deny; discovery of those vendors remains separately gated by compatibility settings.

A symlinked $GROK_HOME or a hooks-paths entry with a symlink component is refused at sandbox start (prevents retargeting). Existing parent directories of protected paths are pinned so they cannot be renamed out from under the deny (siblings remain writable). On Linux, nested user namespaces are disabled inside bubblewrap so mount binds cannot be rearranged. Project hooks remain gated by folder trust. The devbox profile does not apply this protection (disposable VMs). Profiles that require it refuse to start if the kernel policy cannot be applied (including Linux without verified read-only mounts).


Custom Profiles

Create custom sandbox profiles in ~/.grok/sandbox.toml (global) or .grok/sandbox.toml (per-project):

[profiles.project]
# Start from a built-in profile, then add overrides
extends = "workspace"
restrict_network = true

# Paths the agent can read but NOT write/delete
read_only = ["/data"]

# Additional writable paths
read_write = ["/tmp/scratch"]

# Paths or globs to kernel-deny (read + write/rename, enforced; see notes below)
deny = ["/data/shared-secrets", "**/.env", "**/*.pem"]

Use the custom profile:

grok --sandbox project

A custom profile can't reuse a built-in name. --sandbox devbox always runs the built-in devbox profile, shadowing any [profiles.devbox] you define.

If the user and project files define the same custom profile differently, Grok uses the user profile and shows a startup warning. Run /doctor to see both file locations and how to resolve the conflict. Identical definitions do not produce a warning.

Custom Profile Fields

Field Type Description
extends String Base built-in profile to inherit from (workspace, devbox, read-only, strict). Defaults to workspace when omitted
restrict_network Boolean Block network access for child processes
read_only String[] Additional read-only paths
read_write String[] Additional read-write paths
deny String[] Paths or globs to kernel-deny (read + write/rename; see notes). An entry with *, ?, or [ is a glob

Note on deny: A non-empty deny list is kernel-enforced. Denied paths are read-denied and write/rename-denied via Seatbelt on macOS and a bwrap bind-over on Linux, so a denied path can neither be read (via bash, grep, or subagents) nor relocated out of the deny set and read elsewhere (the mv secret x && cat x bypass is closed). On Linux, read-deny requires bubblewrap: if it is missing (or any single deny path can't be bound), Grok refuses to start rather than run with denied paths exposed (devbox, which only write-denies /data, still falls back to Landlock). Writes to paths not in deny are controlled by what you grant in read_write.

Globs in deny: An entry is a glob if it contains *, ?, or [. Those characters always mean glob — to deny a literal file whose name contains them, name a parent directory instead. The supported, gitignore-style subset is:

  • * — any run of characters within one path segment (stops at /)
  • ? — exactly one character within a segment
  • ** — spans directories (as a whole path segment, e.g. **/, a/**); **/ also matches zero directories, so **/.env matches .env and sub/.env
  • [abc] / [a-z] — character classes; a leading ! or ^ negates ([!a] and [^a] both mean "not a")

Brace alternation ({a,b}), backslash-escapes, and the unusual class forms []…] (literal ] first) and POSIX [[:…:]] are not supported, so the two platforms can never interpret a glob differently. A glob using an unsupported metacharacter, or one that is malformed, makes Grok refuse to start (fail closed) on both platforms — write *.pem and *.key as separate entries rather than *.{pem,key}.

Relative globs are anchored at the workspace; absolute globs (e.g. /home/**/.ssh) at their literal prefix. Non-glob entries keep exact-path matching. Enforcement otherwise differs by platform:

  • macOS is airtight: each glob becomes a Seatbelt regex applied at runtime, so matching files are denied even if created after Grok starts.
  • Linux is best-effort: a mount namespace can't glob at runtime, so each glob is expanded to the files that exist at launch and those are bound over. Files created later that match a glob are not covered — name exact paths for anything that must be airtight on Linux. A glob that matches too many files, or whose tree is too deep/broad to walk, makes Grok refuse to start rather than under-enforce.

How It Works

The sandbox is applied to the entire grok process at startup using kernel primitives -- not per-command wrapping. This means all tool operations are covered:

  • read_file, search_replace, list_dir -- restricted by Landlock/Seatbelt in-process
  • bash commands, grep (rg) -- child processes inherit FS restrictions automatically
  • Network -- on Linux, child processes can be blocked via seccomp; on macOS this is a no-op

The sandbox is irreversible once applied. The agent cannot relax restrictions at runtime.


Resuming Sessions

The profile a session was started with is saved with the session and is fixed for the life of the session. When you resume it (grok --resume <id>, grok --continue, or grok -r), Grok restores that same profile automatically — so a session started with --sandbox workspace won't silently come back under a stricter default and break commands that previously worked.

Resuming will not change a session's sandbox:

  • Omitting --sandbox on resume uses the session's saved profile.
  • Passing --sandbox <profile> that matches the saved profile is allowed.
  • Passing --sandbox <profile> that differs from the saved profile is refused with an error — changing a resumed session's sandbox is a safety footgun (it could widen access the session was meant to be confined to, or break a session that relied on broader access). Start a new session to use a different profile.

Profile resolution order for a new session:

  1. An explicit --sandbox <profile> flag or GROK_SANDBOX environment variable
  2. The [sandbox] profile in your config
  3. off (no sandbox)

Platform Support

Platform Mechanism Minimum Version
Linux Landlock Kernel 5.13 or later
macOS Seatbelt macOS (all versions)

If the sandbox cannot be applied (e.g., unsupported kernel, missing entitlements), Grok logs a warning and continues without enforcement. The exception is an explicitly-requested custom profile: on both macOS and Linux, if it cannot be applied (unknown profile, malformed sandbox.toml, or — on Linux — bubblewrap unavailable for a non-empty deny), Grok refuses to start rather than run with its denied paths exposed.


Network Restrictions

On Linux, profiles with restrict_network block network access in child processes (bash commands, scripts) via seccomp. On macOS, network blocking is a no-op. Built-in tools that make HTTP requests in-process (web search, LLM API calls) are never affected -- the agent needs network access to function.

In practice, on Linux this means:

  • web_search, web_fetch, and the LLM API always have network access
  • bash commands like curl, wget, and npm install are blocked when restrict_network is enabled

Shell Environment Policy

The sandbox controls which files and network a subprocess can reach. The top-level [shell_environment_policy] table controls which environment variables it inherits, so a tool command the model runs cannot read a secret that happens to sit in your shell environment.

[shell_environment_policy]
inherit = "core"                 # all (default) | core | none
ignore_default_excludes = false  # also drop *KEY* / *SECRET* / *TOKEN*
exclude = ["ACME_*", "CI_*"]     # drop these names
include_only = ["PATH", "HOME"]  # if set, keep only these names
set = { MY_FLAG = "1" }          # force these values

Grok builds the child environment in order: it starts from inherit (all keeps everything, core keeps a small platform set such as PATH and HOME, none starts empty); drops the built-in secret patterns *KEY*, *SECRET*, and *TOKEN* unless ignore_default_excludes = true; drops any exclude matches; applies set; and, when include_only is non-empty, keeps only the matching names. Patterns are case-insensitive globs (*, ?).

The default (inherit = "all", ignore_default_excludes = true) leaves the environment untouched, so nothing changes until you configure a policy. On the non-persistent backend the policy also filters variables captured from your login shell, so an .rc file export cannot slip a secret past exclude or include_only. The persistent shell is one exception: it applies the policy to its base environment, but variables that an .rc file exports during login are replayed from a snapshot and are not re-filtered, so keep secrets out of shell startup files there. Enforcement covers the bash tool and terminals on macOS, Linux, and Windows.


Event Logging

Sandbox events are logged to ~/.grok/sandbox-events.jsonl for debugging. Events include:

  • Profile applied (which profile, timestamp)
  • Violations (attempted access to denied paths)

When to Use Sandbox Mode

Use workspace when:

  • Working on your own projects and you want basic write protection
  • Running in shared environments where you want to limit the scope of changes

Define a custom profile with a deny list when:

  • You need to block specific files (e.g. .env or credential paths) on top of a base profile
  • You need kernel enforcement that covers bash, grep, and subagents — not just the read_file tool

Use read-only when:

  • Reviewing code you do not trust
  • Exploring a codebase without risk of accidental modification
  • Running code analysis or audits

Use strict when:

  • Analyzing untrusted or third-party code
  • Running in security-sensitive environments
  • You want maximum isolation

Skip sandbox when:

  • The agent needs to install dependencies (npm install, pip install)
  • The agent needs to modify files outside the working directory
  • You are working in a trusted environment and want maximum flexibility

Trade-offs

Aspect Without Sandbox With Sandbox
Safety Agent has full system access Agent restricted to profile rules
Capability Can do anything Limited by profile
Performance No overhead Negligible overhead
Recovery Must trust the agent Kernel enforces boundaries

The sandbox enforces limits at the OS level -- through Landlock or a mount namespace on Linux, and Seatbelt on macOS -- not a separate VM.