Skip to content

LordOS — a free, fast and secure GNU/Linux distribution based on Debian, with its own desktop environment: display server and Wayland compositor in C, a React UI engine on QuickJS, and private native apps.

21.2 KB
Raw Blame

Security review before 0.8.0-beta

An adversarial review of everything that changed since the 0.7.0 release commit (389abc94e0821d2f1f7e611fc6ceed05ea8843c0..847914f, about 40 commits). It was done on 2026-10-18 and checked on VM 118.

The attackers considered:

  • a user with a normal account;
  • a malicious .lpk package;
  • a malicious Linux app or .desktop file, restricted with any of the firewall's restrictions;
  • crafted DNS answers;
  • a hostile release mirror.

Every finding is listed below with its status. Findings rated high or medium are fixed; the low-rated ones are fixed where that was cheap and listed otherwise.

Tests:

  • make -C server test stays green with 0 warnings;
  • npverify-test loads every BPF program into the running 6.12 kernel (38/38);
  • the end-to-end scripts in server/tests/fwlinux/ are listed per finding.

Severity: critical, high, medium, low, info.

Findings

# Severity Area Finding Status Commit
1 High Launcher (apps.c, sec/sandbox.c) The working directory was entered before the tmpfs went over the home and runtime folders, so it still pointed at the real ones. A .desktop with Path=/run/user/<uid> reached the session bus by a relative connect("bus"), then systemd-run --user ran anything outside the sandbox. Without Path=, sockets in the home were reachable (ssh ControlMaster, gpg-agent). Reproduced on VM 118 against 0.7.0+main. Fixed: the directory is entered again inside the view (cfg.cwd, else the home). Test: files.sh cwd_relative_bus. 594399c
2 High Launcher, cgroups app_scope_enter() returned void, and the launcher ran the app even when placing it in its scope failed. That happened for ids with + or over about 62 characters, ids with .., and when lord-sessiond was unreachable. The app then ran in lord-server's own leaf as lordos: no network rules, no device, Bluetooth, PipeWire or notification restrictions. An lpk-<x> desktop id also took the lpk:<x> identity. Fixed: desktop ids are limited to [A-Za-z0-9._-], at most 62 characters, no .. and no lpk- prefix. app_scope_enter returns int, and inside a LordOS session a failed placement never execs (_exit(126)). 594399c
3 High PipeWire guard (sec/pwguard.c) It failed open in three cases: a full client table (256) left a new client untracked, so its stream was never destroyed; a full node table (1024) let a stream go unseen; and a permission list over 16 KB was logged and never sent, leaving the client every permission. Fixed: a client the guard can't follow is disconnected, an unstorable capture stream or node of a restricted or unknown owner is destroyed, and permission lists go in chunks of 400 entries (PipeWire merges them). c1f12c7
4 High PipeWire guard and pipewire-pulse A restricted app (no-mic, even with no-files) could load module-loopback from the microphone into a pipe or a TCP server through pipewire-pulse. The module's stream belongs to the pipewire-pulse daemon, which counted as unrestricted. Reproduced on VM 118. Modules could also open network connections (tunnel, RTP) or create a FIFO at any path of the home. Fixed: module loading is off (/etc/pipewire/pipewire-pulse.conf.d/lordos-no-modules.conf, shipped by the package and the ISO). The daemon's own capture streams are closed while a restricted app runs. Test: mic.sh 8. c1f12c7, 81e1398
5 High Session-side restrictions kernel.yama.ptrace_scope was 0 (Debian's default), so a Linux app restricted without no-files could ptrace lord-server, PipeWire or another app of the same user and act through it, defeating every user-space restriction and the network rules. Fixed: the package ships /usr/lib/sysctl.d/50-lordos-ptrace.conf (kernel.yama.ptrace_scope = 1) and applies it in postinst. Only descendants can be traced, and /proc reads (Monitor, the guard) are unaffected. fb02854
6 Medium Launcher An "Open with" folder grant could include protected folders below it. Opening ~/.config or ~/.local/share read-write handed the app autostart/, firewall.json, applications/ and dbus-1/. Reproduced on VM 118. Fixed: no writable grant on the home itself, or on a folder that contains or lies inside .ssh, .gnupg, systemd, dbus-1, environment.d, user-tmpfiles.d, ~/.local/bin, ~/bin, autostart, applications or the lordos folders. More startup files are added (.bash_aliases, .bash_logout, .zlogout, .xinitrc, .Xresources), and a hard-linked file is granted read-only. Test: files.sh 3b. 594399c
7 Medium Launcher Landlock grants /dev read-write, and /dev/shm and /dev/pts were the session's: other apps' shared memory and the user's terminals (keystrokes, a sudo password). Fixed: a new IPC namespace, a private tmpfs on /dev/shm, and a new devpts instance with /dev/ptmx bound to it. Each fails closed. Test: files.sh read_session_shm, session_pts_hidden. 594399c
8 Medium Launcher Shared caches (fontconfig, Mesa and radv shader caches, the GStreamer registry) were writable. A restricted app could poison what unrestricted apps load as trusted. Fixed: read-only. Test: files.sh write_shared_cache. 594399c
9 Medium Launcher The own-folder deny list missed folders other programs execute from: dbus-1 (service activation), user-tmpfiles.d, terminals', browsers' and editors' config. Fixed: the list is extended. 594399c
10 Medium Launcher Time-of-check to time-of-use on "Open with": the path was checked with realpath() in lord-server but bound (and given to Landlock) later by path, following symlinks. A running instance could swap a component for a symlink to ~/.ssh. Fixed: the opened file is reopened with openat2(RESOLVE_NO_SYMLINKS) and must still be the inode that was checked (dev and inode), both for the bind and for the Landlock rule (also on /media). Otherwise the grant is dropped. 594399c
11 Medium Launcher Without XDG_RUNTIME_DIR no tmpfs went over the runtime dir, so the bus stayed reachable. Fixed: it falls back to /run/user/<uid>, or refuses to launch. 594399c
12 Medium lord-netperm (root) No-mic denied only the ALSA capture minors seen at the last scan (every 5 s). A USB microphone plugged in was open to a restricted app in between, and the app keeps the fd afterwards. Fixed: in the app's own scope every ALSA minor not classified as non-capture is denied (fail closed; np_snd classifies every node). Test: mic.sh 6, with an unknown minor. bd09ae7
13 High lord-netperm, user@ union, launcher The union's no-mic denied the capture nodes to the user's own PipeWire and WirePlumber: while any no-mic app ran, the microphone disappeared for the whole desktop and did not come back (reproduced on VM 118). Taking no-mic out of the union reopened the escape (a no-mic app without no-files could systemd-run --user and open /dev/snd directly). Two exemptions for "the real sound services" were tried and rejected (see "Design change" below). Fixed by design: no process under user@ is exempted for who it is. Every app restricted with no-mic, no-camera, no-usb or no-bluetooth is started in its own view (no session bus, no user manager, Landlock scoping), so it can't escape its scope, and no-mic stays out of the union. Tests: files.sh 3c (for each of the four restrictions: no bus, no systemd/private, no systemd-run, home kept, PipeWire reachable), mic.sh 6 (ALSA capture → EPERM in the scope) and "once lifted, the desktop still records". 00b2c95, see below
14 Medium PipeWire guard A process escaped to user@ (systemd-run --user pw-record) was not restricted. Mitigated: escape is closed for launcher-started apps (#13). While a restricted app runs, the guard also holds user@ clients as restricted, except PipeWire, WirePlumber and pipewire-pulse (exact unit path, alone in it, root-owned binary). That exemption is best effort: it decides only what the guard hides, never a kernel decision. Test: mic.sh 9. c1f12c7, 99a0c7d
15 Medium PipeWire guard (reported by the automated reviewer) The first fix for #14, and the existing pipewire-pulse check, matched the cgroup path by suffix. The user owns user@, so a nested …/x.scope/session.slice/pipewire.service posed as the sound server. Reproduced on VM 118 against the suffix version (it recorded 286 KB). Fixed: exact path built from the uid, sole process in the unit, root-owned executable (same inode). Test: mic.sh 9 (nested fake pipewire and wireplumber units). Other cgroup matching was checked: appid.c (now anchored, #18), netperm scope_app (reads only root-owned directories), cgroup.c (exact strcmp). 99a0c7d
16 Medium PipeWire guard The pulse identity check counted connections but didn't tie them to the claimant. A process in another PID namespace (a no-files app) reports its inner PID, which left slack that a no-mic app could claim with the outer PID. Fixed: a claimed PID must live in our PID namespace (NSpid). Test: mic.sh 7 (claim of a namespaced recorder's outer PID). c1f12c7
17 Medium lord-sessiond (root), pre-existing On a switch the leaving session's framebuffers stayed alive until it handled RP_UNDISPLAY; a stuck or hostile one never does. The next DRM master can read them (GETFB2 + map): the last frame of another user's unlocked desktop. The new keep path made the window longer. Fixed: take_back() removes every framebuffer of the lent file (GETRESOURCES + RMFB) before DROP_MASTER. 913113f
18 Low sec/appid.c The match on /lordos-session- was not anchored. A cgroup the user creates under user@ with the same names posed as an app (attribution only). Fixed: only /system.slice/lordos.service/lordos-session-<own uid>/. c1f12c7
19 Low lord-sessiond Keep (leave the boot splash) could be sent again on a later lend if RP_SHOWN was lost, so the previous picture stayed until the new session drew. Fixed: keep only on the first lend of the process; a failure to create /run/lordos/first-frame is logged. 913113f
20 Low–Medium tools/install/install.sh The verified .deb stayed in a user-owned folder until sudo apt-get install ran. Another process of the user could swap it in between, and root would run its maintainer scripts. Fixed: the package is copied into a root-owned folder, checked again there (SHA-256, package fields) and installed from there. 6e2b66f
21 Low install.sh An empty or non-PEM key override (LORDOS_PUBKEY_FILE) skipped the signature with a warning. Fixed: there is no skip path; a missing or bad key is an error. 6e2b66f
22 Low install.sh An old, validly signed release served as "latest" could downgrade an install. Fixed: refuses to downgrade unless --version is given. A signed version line in SHA256SUMS would be stronger (later). 6e2b66f
23 Low tools/web/deploy.sh The release key was uploaded without being compared with the repository's copy. Fixed. 6e2b66f
24 Low tools/release/sign-release.sh The file list is word-split. Fixed: unexpected file names are refused. 6e2b66f
25 Low hw/notifyd.c If asking the bus who sent a Notify failed, the notification was shown (fail open). Fixed: dropped. c1f12c7
26 Low Website Table alignment used inline style= (dropped by a CSP without unsafe-inline); the team LinkedIn link accepted any scheme. Fixed: classes; https:// only. 0c6a344
27 Low lord-netperm RELOAD can be sent in a loop by any local user (CPU only: 4 connections per uid, each reload rereads a file of up to 4 MB). Open: rate-limit per uid. —
28 Low lord-netperm The hidraw major is read once at startup. If hidraw is a module loaded later, no-usb misses /dev/hidraw*. Debian builds it in (CONFIG_HIDRAW=y). Open. —
29 Low lord-sessiond The session socket's peer isn't checked to be lord-server before a lend. The framebuffer leak it enabled is closed by #17. Open: check SO_PEERCRED pid's /proc/<pid>/exe. —
30 Low notifyd CloseNotification and replaces act on any id, so one app can close or replace another's banner. App name and desktop-entry are the caller's word. Open: record the owner of each id. —
31 Low PipeWire guard The manager socket's peer isn't verified, so an unsandboxed app of the user could replace pipewire-0-manager. This is the same class as "unsandboxed apps can edit firewall.json". Open. —
32 Low Restrictions at startup fwNotifyBlocked in sdk/notify.js returns false on an exception. The rule store is NULL for a moment before js_fw_register. A corrupt firewall.json starts with no rules. Open: fail closed for apps that were restricted. —
33 Low Launcher The abstract unix socket scoping (LANDLOCK_SCOPE_ABSTRACT_UNIX_SOCKET) depends on the build headers. It is present on Debian 13. Open. —
34 Low Launcher /run/dbus (system bus), /run/cups and /run/lordos/*.sock stay reachable from a no-files view. Each was checked: cgroup.sock requires the session leaf and a trusted binary, netperm.sock takes only RELOAD, fand.sock gives STATUS only. Accepted. —
35 Info PipeWire guard A new client has full permissions for milliseconds before the first update_permissions, and a stream can get a few milliseconds of audio before it is destroyed. Documented (FIREWALL.md). —
36 Info Escapes An app not started from the launcher, or already running when it was restricted, may have escaped to user@; the ALSA capture nodes stay open there. An app restricted without no-files can write its own home (startup files, user units, firewall.json), which takes effect later, outside the restriction. Documented (FIREWALL.md); Settings says "restart the app". —
37 High lord-netperm (reported by the automated reviewer) The np_sndsvc exemption (the exact unit cgroup, its sole process's tgid, the packaged binary, no user unit files or drop-ins) was still forgeable. The user controls their own manager: systemd-run --user --unit=pipewire.service passed the file checks (reproduced on VM 118: sndsvc.sh 3 and 4 got the exemption), and so would type-wide service.d drop-ins, set-environment (LD_PRELOAD) and PipeWire's or WirePlumber's own config (modules are code). Removed: replaced by the design in #13. The intermediate commits 30d4e8a and the uncommitted hardening are superseded. see below
38 Medium sec/sandbox.c (reported by the automated reviewer) Landlock's abstract unix socket scoping failed open: with ABI < 6, or built without the scoping constants, the files sandbox went on without it, so a restricted app could reach the session's abstract sockets. Fixed: both the files sandbox and the device view require ABI 6, otherwise the app isn't started. The launcher checks this before forking and reports apps.restrictKernel (a localized notification). The read-only remount of a read-only grant now fails closed too. Test: files.sh 3d (LORD_SANDBOX_MAX_ABI=5, which can only lower the ABI). see below

Design change: no exemption by identity under user@

The user owns their systemd user manager, so anything that runs in user@<uid>.service can be made to look like anything:

  • a nested cgroup with the same names;
  • a process moved into the real unit's cgroup;
  • a unit file, drop-in or type-wide service.d drop-in;
  • a transient unit with the same name (systemd-run --user --unit=pipewire.service);
  • the manager's environment (set-environment, environment.d, LD_PRELOAD);
  • PipeWire's and WirePlumber's own config, which loads modules.

Two attempts to exempt the user's real PipeWire from the union's no-mic were rejected:

  1. matching by cgroup path suffix (#15);
  2. the exact cgroup, its sole process's tgid, the packaged binary's inode and the absence of user unit files (#37).

Both were beaten on VM 118. The design now:

  1. No process under user@ is trusted by lord-netperm for who or what it claims to be.

  2. The escape is closed at the source. A Linux app restricted with no-files, no-mic, no-camera, no-usb or no-bluetooth is started by the launcher in its own view:

    • user, mount, PID and IPC namespaces;
    • a runtime dir with only the Wayland socket, pipewire-0 and pulse/native, so no session bus, no systemd/private and no agents;
    • private /tmp, /var/tmp, /dev/shm and devpts;
    • Landlock scoping of abstract sockets and signals.

    Only no-files also hides the home folder. Such an app can't create or change a unit, or start anything outside its scope, so the device program and the PipeWire guard always find it by its scope.

  3. No-mic stays out of the user@ union. The union keeps covering the network and the other devices for processes that escaped earlier.

  4. The cost: these apps lose the session bus (notifications, portals). A restriction set while the app runs needs a restart to close the escape (the Settings note says so).

Checked and clean

lord-netperm (root):

  • BTF reader (bpfs_btf_func_id):
    • header fields are checked without overflow, and the loop requires pos + 12 <= len;
    • only the 12-byte headers are read, and 12 * vlen can't overflow;
    • string offsets are checked before comparing, and unknown kinds abort;
    • the read is capped at 64 MB, and a big-endian kernel fails safe.
  • /proc/devices and /sys/class/sound: fixed buffers, bounded sscanf (%63s), only the character-device section, NUL-terminated reads, and O_NOFOLLOW on the sound dev files.
  • Device program: only the listed majors, only for an app with that restriction, and never lordos.
  • Bluetooth LSM program: never touches kernel sockets.
  • Host names in notices: they come from the normalized DNS map ([a-z0-9-_.], at most 253 characters). json_quote escapes the quote, the backslash and control characters, and the buffers hold the worst case.
  • firewall.json read: fd-relative, O_NOFOLLOW, an owner check at each step, a regular file of at most 4 MB.
  • RELOAD-only socket: any other frame drops the client. The peer comes from SO_PEERCRED, with a limit per uid.

lord-sessiond:

  • RP_SHOWN (7) collides with no other type; the flags byte is read only when present; mirror_shown goes to seat clients only.
  • /run/lordos/first-frame is created with O_CREAT|O_NOFOLLOW, no truncation, 0644, in a root-owned 0755 folder. A session that fakes RP_SHOWN only ends the splash early.
  • Plymouth script, lordos-splash, and the quit and plymouth-quit units: fixed paths, bounded waits, no user input.

Launcher:

  • Ids: taken only from the file name; never from Exec, Name or StartupWMClass.
  • Namespaces and mounts: MS_REC|MS_PRIVATE; OPEN_TREE_CLONE without AT_RECURSIVE; nosuid and nodev tmpfs; a fresh procfs; identity uid and gid maps with setgroups denied; mounts are locked in the less privileged namespace.
  • Landlock: blocks ptrace of PID 1.
  • Fail closed: every error path refuses to exec.
  • ABI handling: REFER is unhandled under ABI 1 and TRUNCATE is handled from ABI 3.

PipeWire guard:

  • POD parsing (pod_at, pwpod_struct_field, alignment), message size caps, SCM_RIGHTS fds closed, dict_get capped at 4096 entries, and the sock_diag iteration bounded.

notifyd:

  • Only replies from org.freedesktop.DBus are matched, by serial. Pending table of 32, with a 3 s timeout. Bounded markup stripping and JSON.

Broker:

  • restrict.c, brokersrv.c, brokercam.c: the restriction is checked before the permission on every new path, revoking closes running devices, and the id is the verified package id.

js_fw.c:

  • updateRule and ruleApps: every field is validated (app ids, destinations, ports, enums, restriction names), the old rule is copied before any JS getter runs, and frame lengths are bounded. lord.fw exists only in the theme context.

Installer:

  • The signature is checked before anything in SHA256SUMS is used, and HTTPS with TLS 1.2 or later is enforced (redirects too).
  • Quoting is correct, and the GRUB drop-in has a fixed path and fixed content, removed on failure and on uninstall.

Release tooling:

  • No set -x. The private key is never printed or copied, and the signature is verified against the public key.
  • make-sbom.sh uses mktemp with a trap, python3 -I and data in environment variables.
  • build-kernel.sh is unchanged in this range.

Website:

  • md.mjs passes no raw HTML, escapes attributes and allow-lists link schemes.
  • diagram.mjs escapes every field.
  • The FAQ's ld+json escapes <, and the post-build check refuses inline scripts and on*= attributes.
  • site.js uses no innerHTML.
  • deploy.sh is strict and quoted, and leaves the key and install.sh alone on --delete.

Not exercised on the VM

  • The framebuffer removal on a real user switch (#17). The new lord-sessiond was installed on VM 118 and the greeter's first-frame hand-off works ("first frame on the monitor"), but no second user session was switched to.
  • Restarting lordos.service left the old greeter process running until it was killed (systemd: "remains running after unit stopped"). The unit is unchanged in this review, and whether 0.7.0 does the same was not checked. It needs a look before the ISO; the new sessiond lent the monitor only to its own new greeter.