systemd Unit Hardening Check · guide

systemd service hardening: each setting, what it blocks, what it breaks

Who this is for: anyone who has written a .service file to run their own program on a Linux server: a self-hosted app, an API, a worker, a bot, a game server. The typical unit is five lines long, with an ExecStart= and a Restart=always. It runs as root with full access to the filesystem, every device and every system call. If the program has a bug an attacker can reach, they reach the whole machine.

systemd can wall a service off with a few lines in a drop-in, and it costs nothing at run time. The catch is that each setting can break a service that needs what it takes away, so the work is adding settings one at a time and watching the journal.

The order to do it in

  1. Get secrets out of Environment=. The systemd.exec manual says environment variables are not suitable for secrets: a unit's environment is visible to unprivileged users over D-Bus (systemctl show) and is inherited by every child process. Use LoadCredential=name:/etc/app/name and read $CREDENTIALS_DIRECTORY/name, or an EnvironmentFile= owned by root with mode 0600.
  2. Stop running as root. DynamicUser=yes is the shortest route: systemd creates a user for the life of the service and, per the manual, also turns on NoNewPrivileges= and RestrictSUIDSGID=, implies ProtectSystem=strict and ProtectHome=read-only, and gives the service a private /tmp. Add StateDirectory=app for data it keeps. If files must belong to a fixed user, create one and set User=.
  3. Take the binary out of writable places. A unit that runs /tmp/x, or a root unit that runs something under /home, runs whatever is at that path when it starts. Install it under /usr/local/bin or /opt, owned by root.
  4. Add the sandbox with sudo systemctl edit app.service, which writes /etc/systemd/system/app.service.d/override.conf. The checker prints a starter one with only the lines you are missing.
  5. Restart and read the journal (journalctl -u app.service -e). Remove the line that broke it, or loosen it (a writable path, a device, a system call group).

Each setting the checker reads

SettingWhat it blocksWhat it can break
NoNewPrivileges=yesA setuid program (sudo, su, pkexec) raising the service's privileges.Services that really call sudo or a setuid helper. Rarely anything else.
ProtectSystem=strictWrites anywhere except /dev, /proc, /sys and the paths you list. yes and full only cover /usr, /boot and (full) /etc.Anything that writes outside its listed paths. Add ReadWritePaths=, StateDirectory=, CacheDirectory= or LogsDirectory=.
ReadWritePaths=/Nothing: it re-opens the whole filesystem and cancels the line above. The checker flags it as high.—
ProtectHome=yesReading /home, /root and /run/user: SSH keys, shell history, browser profiles.Services that live in or read a home directory. Use read-only, or move the files.
PrivateTmp=yesSharing /tmp with every other process, the usual place for symlink and planted-file attacks.Services that hand files to other programs through /tmp.
PrivateDevices=yesSeeing raw disks, memory and input devices in /dev.GPU, serial, USB or audio access. Use DeviceAllow= for just that device instead.
ProtectKernelTunables=, ProtectKernelModules=, ProtectKernelLogs=, ProtectControlGroups=Changing /proc/sys, loading modules, reading the kernel log, rewriting cgroups.System tools that tune the kernel or manage containers.
SystemCallFilter=@system-service with SystemCallArchitectures=nativeSystem calls a normal service never makes (mount, kexec, reboot, module loading, raw I/O).Unusual programs. Add SystemCallErrorNumber=EPERM so a blocked call fails instead of killing the service while you test.
RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6Raw packet, netlink and Bluetooth sockets.Network tools, VPNs and anything reading interface state over netlink (add AF_NETLINK).
AmbientCapabilities=, CapabilityBoundingSet=The checker flags CAP_SYS_ADMIN, CAP_SYS_MODULE, CAP_SYS_PTRACE, CAP_SYS_RAWIO, CAP_DAC_OVERRIDE, CAP_DAC_READ_SEARCH, CAP_BPF and CAP_SETUID: each is a short step from full root.Keep only what the program needs; a web server on port 443 needs only CAP_NET_BIND_SERVICE.
+ or ! before an Exec*= commandNothing: the prefix runs that command with full privileges outside the sandbox.Keep it to a short fixed helper, never the main process.
UMask=The checker flags a mask that leaves new files writable by other users (0000, 0002 and similar).Use 0027, or 0077 for private data.
ExecStart=/usr/bin/docker run …Little: dockerd starts the container outside the unit, so User= and the sandbox settings confine only the client (podman and nerdctl are treated the same way). The checker says so and skips the missing-sandbox findings.Harden the container itself with the runtime's options (its user, read-only root filesystem, dropped capabilities).
curl … | sh in an Exec*= lineNothing: every start runs whatever that URL serves.Download once, check a hash or signature, install, and run the installed file.

A worked example

A typical self-written unit, and the drop-in the checker writes for it:

[Service]
ExecStart=/usr/local/bin/ledger-api --port 8080
Restart=always
# Starter hardening for ledger-api.service. Install with: sudo systemctl edit ledger-api.service
# Then: sudo systemctl restart ledger-api.service && journalctl -u ledger-api.service -e
# Loosen any line that breaks the service; each is independent.
[Service]
#DynamicUser=yes                                 # run as a throwaway user; check it can still reach its files first
NoNewPrivileges=yes                              # setuid binaries cannot raise privileges
ProtectSystem=strict                             # the whole filesystem read-only; add ReadWritePaths= for what it writes
ProtectHome=yes                                  # home directories hidden
PrivateTmp=yes                                   # a private /tmp
PrivateDevices=yes                               # no hardware devices; remove if it needs a GPU or serial port
ProtectKernelTunables=yes                        # /proc/sys and /sys read-only
ProtectKernelModules=yes                         # no module loading
ProtectKernelLogs=yes                            # no kernel log access
ProtectControlGroups=yes                         # cgroup tree read-only
SystemCallArchitectures=native                   # no foreign-ABI system calls
SystemCallFilter=@system-service                 # the usual service system calls only
RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6 # only Unix and IP sockets

The user line is commented out on purpose: whether the service can run as a throwaway user depends on which files it needs, which only you know.

When to use something else

Open the checker · source (MIT)