There is an edge case where a directory is not readable by the sandbox.
In practice, we've seen very little of it, but it can happen so this
slash command unlocks users when it does.
Future idea is to make this a tool that the agent knows about so it can
be more integrated.
1. Move Windows Sandbox NUX to right after trust directory screen
2. Don't offer read-only as an option in Sandbox NUX.
Elevated/Legacy/Quit
3. Don't allow new untrusted directories. It's trust or quit
4. move experimental sandbox features to `[windows]
sandbox="elevated|unelevatd"`
5. Copy tweaks = elevated -> default, non-elevated -> non-admin
calculated a hashed user ID from either auth user id or API key
Also correctly populates OS.
These will make our metrics more useful and powerful for analysis.
Today, there is a single capability SID that allows the sandbox to write
to
* workspace (cwd)
* tmp directories if enabled
* additional writable roots
This change splits those up, so that each workspace has its own
capability SID, while tmp and additional roots, which are
installation-wide, are still governed by the "generic" capability SID
This isolates workspaces from each other in terms of sandbox write
access.
Also allows us to protect <cwd>/.codex when codex runs in a specific
<cwd>
This fixes a bug where the elevated sandbox setup encrypts sandbox user
passwords as an admin user, but normal command execution attempts to
decrypt them as a different user.
Machine scope allows all users to encyrpt/decrypt
this PR also moves the encrypted file to a different location
.codex/.sandbox-secrets which the sandbox users cannot read.
The elevated setup does not work on non-English windows installs where
Users/Administrators/etc are in different languages. This PR uses the
well-known SIDs instead, which do not vary based on locale
Elevated Sandbox NUX:
* prompt for elevated sandbox setup when agent mode is selected (via
/approvals or at startup)
* prompt for degraded sandbox if elevated setup is declined or fails
* introduce /elevate-sandbox command to upgrade from degraded
experience.
The elevated sandbox creates two new Windows users - CodexSandboxOffline
and CodexSandboxOnline. This is necessary, so this PR does all that it
can to "hide" those users. It uses the registry plus directory flags (on
their home directories) to get them to show up as little as possible.
Never treat .codex or .codex/.sandbox as a workspace root.
Handle write permissions to .codex/.sandbox in a single method so that
the sandbox setup/runner can write logs and other setup files to that
directory.
This is more future-proof if we ever decide to add additional Sandbox
Users for new functionality
This also moves some more user-related code into a new file for code
cleanliness
The elevated setup synchronously applies read/write ACLs to any
workspace roots.
However, until we apply *read* permission to the full path, powershell
cannot use some roots as a cwd as it needs access to all parts of the
path in order to apply it as the working directory for a command.
The solution is, while the async read-ACL part of setup is running, use
a "junction" that lives in C:\Users\CodexSandbox{Offline|Online} that
points to the cwd.
Once the read ACLs are applied, we stop using the junction.
-----
this PR also removes some dead code and overly-verbose logging, and has
some light refactoring to the ACL-related functions
when granting read access to the sandbox user, grant the
codex/command-runner exe directory first so commands can run before the
entire read ACL process is finished.
- Batch read ACL creation for online/offline sandbox user
- creates a new ACL helper process that is long-lived and runs in the
background
- uses a mutex so that only one helper process is running at a time.
The Windows Elevated Sandbox uses two new binaries:
codex-windows-sandbox-setup.exe
codex-command-runner.exe
This PR includes them when installing native deps and packaging for npm
a few fixes based on testing feedback:
* ensure cap_sid file is always written by elevated setup.
* always log to same file whether using elevated sandbox or not
* process potentially slow ACE write operations in parallel
* dedupe write roots so we don't double process any
* don't try to create read/write ACEs on the same directories, due to
race condition
The elevated sandbox ships two exes
* one for elevated setup of the sandbox
* one to actually run commands under the sandbox user.
This PR adds them to the windows signing step
- updating helpers, refactoring some functions that will be used in the
elevated sandbox
- better logging
- better and faster handling of ACL checks/writes
- No functional change—legacy restricted-token sandbox
remains the only path.
the `/approvals` popup fails to recognize that the CLI is in
WorkspaceWrite mode if that policy has extra bits, like `writable_roots`
etc.
This change matches the policy, ignoring additional config aspects.
clean up the code for scanning for world writable directories
One path (selecting a sandbox mode from /approvals) was using an
incorrect method that did not use the new method of creating deny aces
to prevent writing to those directories. Now all paths are the same.
Our Restricted Token contains 3 SIDs (Logon, Everyone, {WorkspaceWrite
Capability || ReadOnly Capability})
because it must include Everyone, that left us vulnerable to directories
that allow writes to Everyone. Even though those directories do not have
ACEs that enable our capability SIDs to write to them, they could still
be written to even in ReadOnly mode, or even in WorkspaceWrite mode if
they are outside of a writable root.
A solution to this is to explicitly add *Deny* ACEs to these
directories, always for the ReadOnly Capability SID, and for the
WorkspaceWrite SID if the directory is outside of a workspace root.
Under a restricted token, Windows always checks Deny ACEs before Allow
ACEs so even though our restricted token would allow a write to these
directories due to the Everyone SID, it fails first because of the Deny
ACE on the capability SID
Fix world-writable audit false positives by expanding generic
permissions with MapGenericMask and then checking only concrete write
bits. The earlier check looked for FILE_GENERIC_WRITE/generic masks
directly, which shares bits with read permissions and could flag an
Everyone read ACE as writable.
The `cap_sid` file contains the IDs of the two custom SIDs that the
Windows sandbox creates/manages to implement read-only and
workspace-write sandbox policies.
It previously lived in `<cwd>/.codex` which means that the sandbox could
write to it, which could degrade the efficacy of the sandbox. This
change moves it to `~/.codex/` (or wherever `CODEX_HOME` points to) so
that it is outside the workspace.
This PR is to unlock future WinGet installation. WinGet struggles to
create command aliases when installing from nested ZIPs on some clients,
so adding raw Windows x64/Arm64 executables lets the manifest use
InstallerType: portable with direct EXEs, which reliably registers the
codex alias. This makes “winget install → codex” work out of the box
without PATH changes.
---------
Co-authored-by: Michael Bolin <mbolin@openai.com>
On Windows, `npm` by itself does not resolve under std::process::Command
which does not consider PATHEXT to resolve it to `npm.cmd` in the PATH.
By running the npm upgrade command via cmd.exe we get proper path
semantics so it actually works.
3 improvements:
1. show up to 3 actual paths that are world-writable
2. do the scan/warning for Read-Only mode too, because it also applies
there
3. remove the "Cancel" option since it doesn't always apply (like on
startup)
1. scan many more directories since it's much faster than the original
implementation
2. limit overall scan time to 2s
3. skip some directories that are noisy - ApplicationData, Installer,
etc.
Show a warning when Auto Sandbox mode becomes enabled, if we detect
Everyone-writable directories, since they cannot be protected by the
current implementation of the Sandbox.
This PR also includes changes to how we detect Everyone-writable to be
*much* faster
- Added the new codex-windows-sandbox crate that builds both a library
entry point (run_windows_sandbox_capture) and a CLI executable to launch
commands inside a Windows restricted-token sandbox, including ACL
management, capability SID provisioning, network lockdown, and output
capture
(windows-sandbox-rs/src/lib.rs:167, windows-sandbox-rs/src/main.rs:54).
- Introduced the experimental WindowsSandbox feature flag and wiring so
Windows builds can opt into the sandbox:
SandboxType::WindowsRestrictedToken, the in-process execution path, and
platform sandbox selection now honor the flag (core/src/features.rs:47,
core/src/config.rs:1224, core/src/safety.rs:19,
core/src/sandboxing/mod.rs:69, core/src/exec.rs:79,
core/src/exec.rs:172).
- Updated workspace metadata to include the new crate and its
Windows-specific dependencies so the core crate can link against it
(codex-rs/
Cargo.toml:91, core/Cargo.toml:86).
- Added a PowerShell bootstrap script that installs the Windows
toolchain, required CLI utilities, and builds the workspace to ease
development
on the platform (scripts/setup-windows.ps1:1).
- Landed a Python smoke-test suite that exercises
read-only/workspace-write policies, ACL behavior, and network denial for
the Windows sandbox
binary (windows-sandbox-rs/sandbox_smoketests.py:1).
Codex isn’t great yet on Windows outside of WSL, and while we’ve merged
https://github.com/openai/codex/pull/4269 to reduce the repetitive
manual approvals on readonly commands, we’ve noticed that users seem to
have more issues with GPT-5-Codex than with GPT-5 on Windows.
This change makes GPT-5 the default for Windows users while we continue
to improve the CLI harness and model for GPT-5-Codex on Windows.
Implement command safety for PowerShell commands on Windows
This change adds a new Windows-specific command-safety module under
`codex-rs/core/src/command_safety/windows_safe_commands.rs` to strictly
sanitise PowerShell invocations. Key points:
- Introduce `is_safe_command_windows()` to only allow explicitly
read-only PowerShell calls.
- Parse and split PowerShell invocations (including inline `-Command`
scripts and pipelines).
- Block unsafe switches (`-File`, `-EncodedCommand`, `-ExecutionPolicy`,
unknown flags, call operators, redirections, separators).
- Whitelist only read-only cmdlets (`Get-ChildItem`, `Get-Content`,
`Select-Object`, etc.), safe Git subcommands (`status`, `log`, `show`,
`diff`, `cat-file`), and ripgrep without unsafe options.
- Add comprehensive unit tests covering allowed and rejected command
patterns (nested calls, side effects, chaining, redirections).
This ensures Codex on Windows can safely execute discover-only
PowerShell workflows without risking destructive operations.
Certain shell commands are potentially dangerous, and we want to check
for them.
Unless the user has explicitly approved a command, we will *always* ask
them for approval
when one of these commands is encountered, regardless of whether they
are in a sandbox, or what their approval policy is.
The first (of probably many) such examples is `git reset --hard`. We
will be conservative and check for any `git reset`
refactors command_safety files into its own package, so we can add
platform-specific ones
Also creates a windows-specific of `is_known_safe_command` that just
returns false always, since that is what happens today.
iceweasel-oaiandGitHub
