Files
pmg/docs/sandbox.md
T
ee684a29a9 feat(sandbox): presets — additive workload allowance bundles (#387)
* feat(sandbox): introduce presets - additive workload allowance bundles

Presets are named, additive-only bundles of sandbox allowances for a
specific workload (git hooks tooling, Astro/Vite/Next.js dev servers).
They solve the per-workload tuning friction from #384 without weakening
the default posture: no built-in profile references a preset, presets
cannot carry deny rules or profile booleans (strict YAML decoding), and
mandatory denies still win everywhere except the existing exact-match
suppression.

- Preset schema with metadata (author, labels) and schema_version gating
- Registry over ordered sources (embedded builtin, user dir); builtin
  wins name collisions; source abstraction is the extension point for a
  future hosted registry and SafeDep cloud sync
- Official presets: git, astro, vite, nextjs (with threat notes)
- Overlay/runtime integration: pmg sandbox allow preset=<name> and
  --sandbox-allow preset=<name>, stored by reference, resolved at apply
  time, missing presets warn (fail closed) instead of aborting
- Profile integration: presets: [...] list resolved after inherits
- CLI: pmg sandbox preset list (metadata filters, --json), show (prints
  YAML with threat notes), lint
- Docs: user guide (docs/sandbox-presets.md) and design spec

Closes #384

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* fix(sandbox): address review findings on presets

- Presets never modify deny lists: a profile authored deny now survives a
  preset allowing the same path (deny-beats-allow keeps it enforced).
  Regression test added.
- Profile inspection commands (show, diff, lint) construct the profile
  registry with the user-aware preset registry so they agree with runtime
  resolution of custom profiles referencing user presets.
- Handle stderr write error when warning about unresolvable presets.
- Compute preset show underline from the uncolored header.
- Use path.Join for embed.FS reads (slash-separated on all platforms).
- Clarify in docs that lint-staged/astro are examples of preset workloads.
- Drop the design spec from the PR per review.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* fix(sandbox): harden preset precedence against authored denies

Addresses external security review findings on the preset mechanism:

- Bubblewrap: a mandatory write-denied path listed in allow_read lost its
  protection when a later writable parent bind covered it (bwrap last
  mount wins) - exactly the git preset shape (allow_read .git/config +
  allow_write .git/**). The mandatory deny now re-binds the path
  read-only after all writable mounts instead of being skipped.
  Regression test asserts mount ordering. Landlock and Seatbelt were
  unaffected (tests added for the same policy shape on Landlock).
- Environment: ScrubEnv is allow-wins, so a preset environment allowance
  could override a profile-authored deny. Preset env allowances
  overlapping an authored deny pattern are now dropped at application
  time (conservative bidirectional glob overlap, fail closed). Surviving
  entries still opt out of built-in credential scrubbing as intended.
- Network: removed allow_outbound from the preset schema. Both platform
  translators are all-or-nothing for outbound (one allow rule means
  blanket network access), so a preset outbound entry would silently
  change network posture far beyond what its YAML conveys. Strict
  decoding rejects the key.
- Added a dual-path expansion equivalence test (profile presets: field
  vs overlay/--sandbox-allow) and documented the precedence guarantees
  in docs/sandbox-presets.md.

Explicit --sandbox-allow and pmg sandbox allow overrides keep their
existing semantics.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* docs(sandbox): document env and preset allowances in allow command and overlay docs

pmg sandbox allow help, the --sandbox-allow flag usage, and the project
overlay docs enumerated only read/write/exec/net types. Add env and
preset to all of them, with an overlay example for persisting an env
allowance and a note on why env entries are not auto-promoted by --last.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* chore(sandbox): trim preset code comments to corner cases and minimal godocs

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* fix(sandbox): exact glob intersection for preset env deny overlap

The bidirectional literal-text heuristic missed overlapping globs with
different literal structure: preset allow AWS_*_KEY and authored deny
AWS_SECRET_* both match AWS_SECRET_ACCESS_KEY but neither pattern
matches the other's text, so the allowance merged and allow-wins
scrubbing exposed the variable. EnvPatternsOverlap now computes exact
intersection non-emptiness for the name glob dialect (case-insensitive,
'*' any sequence, '?' single char) via memoized DP.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* fix(sandbox): preset env allowances are exact names, not globs

Glob-vs-glob intersection is a losing game: every dialect extension
(character classes today) silently reopens the deny-bypass hole.
Restricting preset environment allowances to literal variable names
makes the authored-deny precedence check exact by construction: each
deny pattern is evaluated against the concrete name with the same
matcher ScrubEnv uses at runtime, so the decision cannot diverge from
enforcement regardless of deny dialect. Removes the glob intersection
machinery. Profile and --sandbox-allow env globs are unaffected.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* fix(sandbox): reject mandatory-deny targets in preset paths

Preset validation relied on IsSensitiveProjectTarget, which covers fewer
files than util.DANGEROUS_FILES. A preset naming .git-credentials,
.pgpass, .docker/config.json or .config/gh exactly would exact-match
suppress the mandatory deny; .git/config in allow_write would suppress
the write protection. Preset paths are now checked against
DANGEROUS_FILES (single source of truth), .git/hooks is rejected in any
direction, and .git/config is rejected for write/exec while read stays
allowed for git repo discovery. Docs state the two deliberate opt-outs
precisely.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* feat(sandbox): preset init and edit commands for community authoring

pmg sandbox preset init scaffolds a valid user preset (metadata flags,
threat-note template, starter rule) and refuses built-in names since
builtins win resolution. pmg sandbox preset edit opens the file via the
shared editor package and validates the result, warning when a user
preset is shadowed by a built-in. Docs lead with the scaffolded flow and
spell out builtin-vs-community provenance in preset list.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

* refactor(sandbox): move mandatory-target matching into util

Preset path validation re-encoded knowledge util already owns: the
dangerous-files comparison and hardcoded .git/config and .git/hooks
strings. util now exports GitConfigPath, GitHooksPath (also used by
GetMandatoryDenyPatterns), PathCoveredBy and DangerousFileMatch, and
preset validation consumes them so the mandatory deny policy has a
single definition.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PETCfE4crLcmodosz12qRn

---------

Co-authored-by: Claude <noreply@anthropic.com>
2026-07-21 15:14:07 +05:30

622 lines
27 KiB
Markdown

# Sandbox
PMG sandbox design goal is to protect against unknown supply chain attacks using principle of least privilege.
We do not want to re-invent sandbox and rely on OS native sandbox primitives. This is at the cost of
developer experience, where we have to work within the limitations of the sandbox implementations
that we use.
## Security Model
The sandbox is default-deny. Every operation is blocked unless the policy allows it, and deny rules
win over allow rules when both match. PMG ships mandatory denies for credential files (`.env`,
`.ssh`, `.aws`, `.gcloud` and more). See [dangerous.go](../sandbox/dangerous.go) for the full list.
This protects against accidental credential leaks and some classes of supply chain attacks that
attempt to access credentials. To allow legitimate access, you can opt out of a mandatory deny
via an exact match entry in `allow_read` or `allow_write` (in the policy or via `--sandbox-allow`).
<details>
<summary>Detailed rules</summary>
- **Default deny**: All operations are blocked unless explicitly allowed by the policy. An empty policy grants no access.
- **Deny rules override allow rules**: When a path appears in both allow and deny lists, the deny rule wins. Deny rules are placed after allow rules in the generated sandbox profile to ensure this.
- **Credential and sensitive file protection**: The sandbox blocks read and write access to known list of credential files by default.
- **Git hooks are always blocked**: Write access to `.git/hooks/` in both `$CWD` and `$HOME` is always denied to prevent arbitrary code execution via repository hooks.
- **Git config is blocked by default**: Write access to `.git/config` is denied unless `allow_git_config: true` is set in the policy. This prevents credential helper manipulation.
- **Runtime overrides remove only exact-match deny entries**: When `--sandbox-allow` adds a path to an allow list, only a literal string match in the corresponding deny list is removed. Glob and wildcard deny patterns (e.g., `/etc/**`) are never removed. An exact-match entry in `allow_read` or `allow_write` (policy or runtime) opts out of the mandatory deny for that credential file. `.git/hooks` does not accept opt-outs.
- **Profile inheritance is single-level**: A profile can inherit from one built-in profile. Allow and deny lists are merged using union semantics. Boolean fields (`allow_pty`, `allow_git_config`) in the child override the parent.
- **Variable expansion is runtime-only**: Policy paths use `${HOME}`, `${CWD}`, and `${TMPDIR}` which are expanded when the sandbox is set up, not when the policy is defined.
- **Process-level isolation only**: The sandbox restricts the package manager process and its children. It does not enforce CPU, memory, or disk quotas. Network filtering is coarse-grained — host-level filtering is not enforced on either platform.
</details>
### Mandatory Credential Protection
PMG maintains a known list of credential and sensitive files at
[dangerous.go](../sandbox/dangerous.go) that are blocked by default in the sandbox. PMG injects three deny patterns per file:
- The path under `${CWD}`
- The path under `${HOME}`
- A `**/<file>` glob.
To opt out, list the literal path in `allow_read` or `allow_write` in your policy, or pass
`--sandbox-allow read=...` / `write=...` at runtime. Listing a CWD-absolute or HOME-absolute path
also suppresses the matching `**/<file>` glob on the same direction, so `--sandbox-allow
read=./.env` is enough to read `${CWD}/.env`. Suppression is exact post-expansion match; broad globs
like `${HOME}/**` do not opt out of `${HOME}/.aws`. The unnamed absolute form stays denied.
`.git/hooks` does not accept opt-outs because hooks can execute arbitrary code.
### Environment Variable Protection
Many supply chain attacks steal credentials from the **process environment** rather than from files
(e.g. `AWS_SECRET_ACCESS_KEY`, `GITHUB_TOKEN`, `NPM_TOKEN`, `TWINE_PASSWORD`). When the sandbox is
enabled, PMG scrubs a default-deny list of credential-bearing variables from the package manager
child process before it is spawned. The built-in list is an explicit, curated set of **known** secret
names. See [`DANGEROUS_ENV_VARS`](../sandbox/util/dangerous.go). There are deliberately no generic
`*_TOKEN` / `*_SECRET` catch-alls in the default, because broad wildcards there would risk clipping
legitimate build variables.
Scrubbing is platform-independent (it filters the environment regardless of the OS sandbox driver)
and runs as the last step before launch, after project overlays and `--sandbox-allow` overrides are
merged. Scrubbed variable **names** (never values) are logged at info level. Run with `--debug` to
see what was removed.
The shared base profiles (`npm-restrictive`, `pypi-restrictive`) allow no environment variables.
Each package manager's leaf profile (`npm`, `yarn`, `bun`, `pnpm`, `npx`, `pip`, `pipx`, `uv`,
`uvx`, `poetry`) re-allows only the variables that package manager legitimately needs via an
`environment.allow` block, so package managers keep working:
```yaml
environment:
# Re-permit only what this package manager needs; everything else in the
# default deny list stays scrubbed. allow always wins over deny.
allow:
- NPM_TOKEN
- npm_config_*
# Optionally scrub more than the default. Glob patterns are supported here
# (they are intentionally not in the built-in default).
deny:
- MY_CUSTOM_SECRET
- "*_TOKEN"
```
**Accepted trade-off**: a leaf profile re-allows its own package manager's auth token, so a
malicious package executed during `npm install` can read `NPM_TOKEN`, but not a yarn or bun token, a
PyPI token, AWS key, or other cloud/secret-manager credential, which stay scrubbed. The reverse
holds for the other profiles. This is deliberate: the package manager needs its own auth token to
function.
Configs written before the profile split may still map npm, yarn, or bun to `npm-restrictive`, pnpm
to `pnpm-restrictive`, or pip, pip3, poetry, or uv to `pypi-restrictive`. PMG re-maps these legacy
defaults to the per-package-manager leaf profile at load time, unless a custom policy template
overrides the legacy name, in which case the template wins as before.
Matching is on the variable name, case-insensitive, and supports the same glob syntax as filesystem
rules. A small set of core variables (`PATH`, `HOME`, `LC_*`, `TZ`, ...) is never scrubbed.
## Requirements
- Linux kernel 5.13+ with Landlock enabled (default, no external dependencies)
- Bubblewrap on Linux (fallback for kernels < 5.13, or when `PMG_SANDBOX_DRIVER=bubblewrap` is set)
- Seatbelt on MacOS
<details>
<summary>Bubblewrap Installation on Linux</summary>
For Debian-based Linux distributions, you can install Bubblewrap with the following command:
```bash
sudo apt install bubblewrap
```
For Arch Linux, you can install Bubblewrap with the following command:
```bash
sudo pacman -S bubblewrap
```
For other Linux distributions, you can install Bubblewrap from the package manager of your choice.
See [Bubblewrap Installation](https://github.com/containers/bubblewrap#installation) for more details.
</details>
## Usage
- Make sure sandbox is enabled in your `config.yml` file. See [configuration](./config.md) for the configuration schema.
- Make sure sandbox profiles are configured for the package managers you want to sandbox.
See [configuration](./config.md) and [config/config.template.yml](../config/config.template.yml) for the configuration schema.
Once sandbox is enabled, you can run package manager commands with sandbox protection.
Run `pmg sandbox doctor` to see platform specific sandbox setup and driver status. Continue using
PMG as usual, sandbox will be applied to configured package managers automatically.
```bash
pmg npm install express
```
Explicitly enable sandbox if not enabled in the `config.yml` file:
```bash
pmg --sandbox --sandbox-profile=npm-restrictive npm install express
```
Run sandbox with custom policy file:
```bash
pmg --sandbox --sandbox-profile=/path/to/custom-policy.yml npm install express
```
### Sandbox Profile Commands
Use profile commands to inspect, create, and validate sandbox profiles.
```bash
# List built-in and user profiles
pmg sandbox profile list
# Scaffold a user profile that inherits from a built-in
pmg sandbox profile init my-npm --from npm-restrictive
# Open a user profile in $VISUAL / $EDITOR, then validate and lint it
pmg sandbox profile edit my-npm
# Show a profile, or its fully resolved policy
pmg sandbox profile show npm-restrictive
pmg sandbox profile show npm-restrictive --resolved
# Lint a built-in, user profile, or profile file
pmg sandbox profile lint npm-restrictive
pmg sandbox profile lint ./my-profile.yml --strict
# Compare two resolved profiles
pmg sandbox profile diff npm-restrictive pypi-restrictive
```
`pmg sandbox profile edit` resolves the profile name to its file under the user profile
directory, opens it in your editor, and re-validates and lints the result when the editor
exits. Built-in profiles are embedded in the binary and cannot be edited; create an editable
copy with `pmg sandbox profile init <new-name> --from <builtin>`.
To activate a user profile for a package manager, reference it by name in `config.yml` — no
policy template is needed for profiles in the user profile directory:
```yaml
policies:
pnpm:
enabled: true
profile: my-pnpm
```
Policy templates (see below) are only needed to override a built-in profile name or to load a
policy file from outside the user profile directory.
### Sandbox Debug Commands
Use these commands when a sandboxed package manager command fails and you need to inspect why.
```bash
# Check sandbox driver availability and host setup
pmg sandbox doctor
# List recent sandbox denials captured by PMG
pmg sandbox violations list
# Explain the latest captured denial and show a suggested override when possible
pmg sandbox explain --last
# Inspect the resolved policy PMG will apply
pmg sandbox profile show npm-restrictive --resolved
```
`pmg sandbox doctor` runs platform-specific checks for the current host. Cached violation reports
used by `violations list` and `explain --last` are currently produced by macOS Seatbelt diagnostics;
on Linux, Bubblewrap and Landlock denials may only appear as command errors such as `EACCES`.
### Runtime Allow Overrides
Use `--sandbox-allow` to make one-off exceptions without creating a custom profile. This is useful
when a command needs access that the default profile blocks.
```bash
# Allow writing to a specific file
pmg --sandbox-allow write=./.gitignore npx create-next-app@latest
# Allow executing a binary blocked by the profile
pmg --sandbox-allow exec=$(which curl) npm install some-package
# Allow outbound connection to a private registry
pmg --sandbox-allow net-connect=npm.internal.corp:443 npm install @corp/private-pkg
# Allow a dev server to bind to a local port
pmg --sandbox-allow net-bind=127.0.0.1:3000 npx some-dev-tool
# Re-allow a sensitive environment variable that the profile scrubs by default
pmg --sandbox-allow env=AWS_PROFILE aws-cdk-using-package install
# Multiple overrides
pmg \
--sandbox-allow write=./.gitignore \
--sandbox-allow exec=$(which curl) \
npm install some-package
```
Supported types: `read`, `write`, `exec`, `net-connect`, `net-bind`, `env`.
For `env`, the value is an environment variable **name** or name glob (e.g. `NPM_TOKEN`,
`npm_config_*`) and is kept verbatim. It is not path-resolved. It is an allow-only override that
re-permits a variable the profile would otherwise scrub; allow always wins, so there is no deny list
to edit. If a command fails with an auth error, re-run with `--debug` and look for a "scrubbed" log
line naming the variable, then re-allow it with `--sandbox-allow env=NAME`.
Overrides are non-persistent (apply to current invocation only) and logged in the event log for
auditing. An override adds the path to the allow list and removes an exact match entry from the
corresponding deny list. Glob deny patterns are never removed. An exact-match entry in `allow_read`
or `allow_write` (in the policy or via `--sandbox-allow read=...` / `write=...`) opts out of the
mandatory deny for that credential file. PMG treats both channels as explicit user intent.
Suppression is exact-match only; broad paths or globs do not opt out. `.git/hooks` does not accept
opt-outs.
### Project Overlays
A project overlay is a per-repository allow-set that PMG applies automatically on every run in
that repo, so you do not have to retype the same `--sandbox-allow` flags. The overlay is local
only, keyed by the repo's git toplevel (CWD when not a git repo), and stored under
`<config_dir>/sandbox/overlays/`. It is purely additive on top of the resolved policy, never
weakens mandatory denies, and is ignored when `global_lockdown` is set.
```bash
# Save manual allowances for the current repo
pmg sandbox allow write=./.astro net-bind=localhost:4321
# Persist an environment variable allowance so the profile stops scrubbing it
# in this repo (same semantics as --sandbox-allow env=..., but saved)
pmg sandbox allow env=AWS_PROFILE
# Promote the primary violation from the most recent cached report
pmg sandbox allow --last
# Promote every safe FS/exec violation from that report
pmg sandbox allow --last --all
# Allow a sensitive path (e.g. .env, .npmrc) explicitly
pmg sandbox allow --force read=./.env
# Inspect the current repo's overlay
pmg sandbox project show
pmg sandbox project show --json
# List overlays across all known repos
pmg sandbox project list
# Delete the current repo's overlay
pmg sandbox project reset --yes
```
### Presets
A preset is a named, additive-only bundle of allowances for one workload (git hooks tooling,
an Astro/Vite/Next.js dev server, ...). Instead of discovering allowances one denial at a
time, apply a curated bundle:
```bash
pmg sandbox preset list
pmg sandbox preset show git
pmg sandbox allow preset=git preset=astro
```
Presets can also be attached to a profile via a `presets:` list. See
[sandbox-presets.md](sandbox-presets.md) for usage and how to author your own.
Notes:
- `--last`/`--last --all` only auto-promotes filesystem and exec denials. Network allowances
(`net-connect`, `net-bind`) and environment allowances (`env`) must be passed manually as
`type=value` because drivers do not classify network denials yet and environment scrubbing is
logged rather than recorded as a violation. Run with `--debug` to see scrubbed variable names,
then persist with `pmg sandbox allow env=NAME`.
- `pmg sandbox allow` refuses sensitive targets (`.env*`, `.npmrc`, `.ssh`, `.aws`, `.kube`,
`.gnupg`, ...) unless `--force` is given.
- Applied overlay entries are recorded in the audit event log with a `+overlay` source tag so
they can be distinguished from `--sandbox-allow` flags.
<details>
<summary>Custom policy overrides using Policy Templates</summary>
Policy templates allow custom policy overrides. To setup custom policy overrides for your package manager,
start by looking up the PMG configuration directory:
```bash
pmg setup info
```
Create a new policy template file in the PMG configuration directory and edit it to suit your needs:
```bash
# Set the PMG configuration directory
export PMG_CONFIG_DIR="/path/to/pmg/config/dir"
# Create the policy template file
cat > $PMG_CONFIG_DIR/sandbox-custom-policy.yml <<EOF
name: pnpm-macos-custom-sandbox
description: Custom profile for pnpm in MacOS
inherits: npm-restrictive
package_managers:
- pnpm
allow_pty: true
filesystem:
allow_write:
# pnpm i need write access here
- ${HOME}/Library/pnpm/.tools/**
# pnpm i creates these tmp files in local dir, at least on MacOS
- ${CWD}/_tmp_*
# pnpm self-update (or likely update) creates temporary package.json files
# for writing. This is likely for atomic update using filesystem rename operation
# which guarantees atomicity
- ${CWD}/package.json.*
# Need access for dependency resolution
- ${CWD}/.pnpm-store
# Additional deny rules for extra security
deny_write:
- ${CWD}/.env
- ${CWD}/.env.*
EOF
```
Edit PMG configuration file to use the custom policy template and override the default
policy for your package manager:
```yaml
policy_templates:
pnpm-macos-custom-sandbox:
path: ./sandbox-custom-policy.yml
policies:
pnpm:
enabled: true
profile: pnpm-macos-custom-sandbox
```
Next time you run `pmg pnpm install`, the custom policy template will be used instead of the default policy.
</details>
## Supported Platforms
| Platform | Supported | Implementation |
| -------- | --------- | ----------------------------------- |
| MacOS | Yes | Seatbelt sandbox-exec |
| Linux | Yes | Landlock (default, kernel 5.13+) or Bubblewrap (fallback) |
| Windows | No | Not yet supported |
### Platform-Specific Limitations
<details>
<summary>Linux (Landlock, default)</summary>
**Default sandbox on kernel 5.13+**: Landlock provides kernel-native filesystem access control
without requiring external binaries or unprivileged user namespaces.
For the architecture, design tradeoffs, and known limitations see
[sandbox-landlock.md](./sandbox-landlock.md).
**Deny enforcement**: Deny rules (DenyRead, DenyWrite, DenyExec) are enforced via seccomp
user notifications. This introduces a small TOCTOU window (microseconds) between reading
the path and responding.
**Deny enforcement across the process tree**: seccomp-notify resolves the path argument of
an intercepted `openat(2)` by reading `/proc/<pid>/mem` of the trapping process. PMG ships
this in a two-stage architecture so enforcement applies to direct targets AND every
descendant (grandchildren, great-grandchildren, etc.):
1. The helper process (`pmg __landlock_sandbox_exec`) clones a tiny shim
(`pmg __landlock_shim`) with `CLONE_NEWUSER` and a uid/gid map of `0 -> host uid`.
The shim runs as uid 0 inside a fresh user namespace so it has `CAP_SYS_ADMIN` in that
namespace.
2. The shim installs the seccomp-notify filter **without** `PR_SET_NO_NEW_PRIVS` (permitted
by `CAP_SYS_ADMIN` in the ns). It then applies Landlock and `execve`s the real target.
3. Because `NO_NEW_PRIVS` was never set, subsequent `execve` calls in the tree do **not**
reset `dumpable` to 0, so the helper can keep opening `/proc/<pid>/mem` for any
descendant. Deny rules like `~/.ssh` are enforced for the full process tree.
The user namespace is purely a capability vehicle. Host uid/gid are preserved through the
mapping, so targets see the same filesystem ownership they normally would. Tools that
refuse to run as root (npm's root-in-container warning) are unaffected because the
outside-view uid never changes.
**Requirements**: unprivileged user namespaces must be enabled (`unprivileged_userns_clone=1`
on Debian/Ubuntu; default on most modern distros). If disabled, the helper fails with an
EPERM on `clone()` and the sandbox falls back to Bubblewrap.
**Network filtering**: Not enforced. Landlock supports TCP port filtering only (V4+, no hostname).
Use `--proxy-mode` for network control.
**PID/IPC namespace isolation**: Applied best-effort via `CLONE_NEWPID|CLONE_NEWIPC|CLONE_NEWNS`.
If unavailable, a warning is printed and the command continues. Set `PMG_SANDBOX_DRIVER=bubblewrap`
to force Bubblewrap if namespace isolation is required.
**`/proc` access**: The sandbox supervisor requires `/proc` read access. When PID namespace
isolation succeeds, `/proc` is scoped to the child's namespace. When it fails, `/proc`
exposes all system processes.
**Fallback**: If Landlock is unavailable (kernel < 5.13), Bubblewrap is used automatically.
Set `PMG_SANDBOX_DRIVER=bubblewrap` to force Bubblewrap.
</details>
<details>
<summary>Linux (Bubblewrap, fallback)</summary>
**Filesystem permissions are coarse-grained**: [Bubblewrap](https://github.com/containers/bubblewrap) uses bind mounts for filesystem isolation.
To prevent `Argument list too long` errors with large directory trees, PMG automatically uses
coarse-grained fallback strategies when glob patterns match many files.
**Fallback Behavior:**
- **Small patterns** (< 100 matches): Individual files are mounted (fine-grained, most precise)
- **Large patterns** (> 100 matches): Parent directory is mounted (coarse-grained, scalable)
- **Threshold**: 100 paths per pattern triggers coarse-grained fallback
**Network filtering**: All-or-nothing network isolation (via `--unshare-net`). Host-specific
filtering is not enforced.
**Per-direction mandatory deny is asymmetric on Linux**: bwrap has no primitive that allows writes
while denying reads for the same path. `--bind` exposes both directions; `--tmpfs` and
`--ro-bind /dev/null` block both. If you opt out of write for a mandatory deny path (e.g., list it
in `allow_write` but not `allow_read`), the bind mount also exposes reads, and PMG cannot enforce
the read-side mandatory deny. PMG warns via `log.Warnf` when it detects this case. macOS Seatbelt
does not have this limitation; its `file-read*` and `file-write*` rules are independent.
</details>
<details>
<summary>macOS (Seatbelt)</summary>
**Network lockdown (`network_via_proxy_only`)**: Fine-grained `host:port` filtering is not
expressible in Seatbelt, so per-host control happens at the PMG proxy instead. With
`network_via_proxy_only: true`, the sandbox denies all non-loopback outbound network; only the
running PMG proxy's port is reachable, and profiles with `allow_network_bind` additionally keep
loopback↔loopback connects open. Raw sockets, QUIC, and arbitrary non-loopback ports are blocked
at the kernel.
Direct DNS is disabled by default (the proxy resolves names); `allow_direct_dns: true` re-opens
it. The Go profile ships with lockdown enabled.
Lockdown is fail-closed: it requires the proxy flow, and pmg errors out rather than running
without confinement when no proxy is available — including on Linux, where
`network_via_proxy_only` is not yet supported (the drivers reject it with a clear error, never a
silent fallback).
For local development: plain commands like `npm run dev` are not sandboxed unless
`enforce_always` is set; loopback↔loopback traffic keeps working via `allow_network_bind` (as
noted above); and
proxy-honoring clients reach any destination through the proxy CONNECT path. What breaks under
lockdown is direct non-loopback sockets — tools that ignore `HTTP_PROXY`/`HTTPS_PROXY` and
non-HTTP wire protocols (e.g. Postgres or Redis clients pointed at non-loopback hosts). Such
denials render as:
> direct network access blocked by network_via_proxy_only — traffic must flow through the PMG proxy (a tool may have ignored HTTP_PROXY/HTTPS_PROXY)
</details>
## Concepts
PMG layers three concepts: a **Policy** is the rule set defining what is allowed and denied. A
**Profile** is a named binding from a package manager to a policy. A **Policy Template** maps a
profile name to a YAML file so you can override the built-ins.
<details>
<summary>Detailed concepts</summary>
### Policy
Policy is a set of rules that define the allowed and denied actions for a package manager. A sandbox implementation, such as
`sandbox-exec` on MacOS enforces the policy.
PMG defines its own policy model. The design goal is simplicity and ease of use. Sandbox implementations are expected to translate
the policy model into their own native policy format. Rules for policy are:
- Deny by default unless explicitly allowed
- Deny rules have higher priority than allow rules
- Policy profile allows binding package managers to a specific sandbox policy
- Package manager must have a sandbox profile when sandbox is enabled
- Package manager specific sandbox profile may be disabled to skip sandbox for the package manager
### Profile
Profile is a named reference to a policy. It is used to associate a policy with a package manager. PMG ships with a set of built-in profiles
that are used to enforce the policies for the package manager. See [sandbox/profiles](../sandbox/profiles) for the list of built-in profiles.
Custom profiles can be created by copying a built-in profile and modifying the rules to suit the needs.
See [sandbox/profiles/README.md](../sandbox/profiles/README.md) for more details.
### Policy Template
Policy template is a configuration primitive for overriding a built-in profile or creating a custom profile. It is used to map a profile name to a path.
See [config/config.template.yml](../config/config.template.yml) for an example.
</details>
## Threat Model
PMG trusts policy files and the operator's CLI as the source of intent. The sandbox implementation
enforces what the policy declares. Translation from PMG's YAML to the native sandbox format must
not weaken it. Variable interpolation consumes only trusted sources.
<details>
<summary>Detailed assumptions</summary>
- Policy files are trusted
- Policy enforcement is a sandbox implementation concern
- YAML to sandbox specific policy translation must not make the policy weaker than the original policy
- Variable interpolation in policy files must consider only trusted sources
</details>
## Enforcement
The sandbox implementation currently only support `block` mode. This means, any policy violation will block the execution of the
package manager command.
## Debug
### MacOS
OSX sandbox implementation is based on [Chromium OSX Sandbox Design](https://www.chromium.org/developers/design-documents/sandbox/osx-sandboxing-design/)
and [Anthropic Sandbox Runtime](https://github.com/anthropic-experimental/sandbox-runtime). Current implementation does not support
identifying sandbox policy violations.
To manually investigate sandbox policy violations, you can use the following command:
```bash
APP_LOG_LEVEL=debug APP_LOG_FILE=/tmp/pmg-debug.log pmg --sandbox --sandbox-profile=npm-restrictive npm install express
```
Find the log tag in the debug log file and use it to investigate the sandbox policy violation.
```bash
grep "PMG_SBX_" /tmp/pmg-debug.log
```
Use `log(1)` to filter the log file by the log tag.
```bash
log show --last 5m --predicate 'message ENDSWITH "PMG_SBX_${TAG}"' --style compact
```
### Linux
Linux sandbox implementation uses Bubblewrap for namespace-based isolation. Enable debug logging to see translated sandbox arguments:
```bash
APP_LOG_LEVEL=debug APP_LOG_FILE=/tmp/pmg-debug.log pmg --sandbox --sandbox-profile=npm-restrictive npm install express
```
Review the debug log to see the translated `bwrap` command-line arguments:
```bash
grep "Bubblewrap arguments" /tmp/pmg-debug.log
```
To debug sandbox violations, you can manually test commands with increased verbosity by running the sandbox command directly:
```bash
# Extract the bwrap command from debug logs and run with --verbose
bwrap --verbose [arguments...] -- npm install express
```
**Note**: Unlike macOS, Bubblewrap does not provide real-time violation logging. Policy violations typically manifest as `EACCES` (Permission denied) errors.
## References
- <https://github.com/anthropic-experimental/sandbox-runtime>
- <https://geminicli.com/docs/cli/sandbox/>
- <https://github.com/containers/bubblewrap>