Community packages (AUR): checking and sandbox

lux install <name> uses the community source (AUR) only when the name is not in the LuxOS and Arch Linux repositories, the LuxOS .lux apps or Flathub (order in package-management.md). AUR build scripts are written by other users and are not checked by LuxOS, so lux checks them itself and builds them isolated by default.

Checking, sandbox and system installs

  • Before building, the PKGBUILD, .install files and every other file in the repository are checked in two layers: 1. Fixed rules (offline, deterministic): downloading and running code (curl | sh, download straight into $pkgdir, download to /tmp + chmod +x), hidden payloads (base64/hex, text reversed with rev, eval of fetched text), writes outside $pkgdir, touching accounts/sudo/SSH, persistence (.bashrc, per-user autostart/systemd unit), /etc/hosts, resolv.conf, ld.so.preload, sudoers/SUID, disabling the firewall or a security service, replacing a system binary, data upload / credential, cookie or SSH-key theft, cryptominer names, sources from a bare IP or look-alike host, missing checksums (SKIP) on a remote download, symbolic links pointing outside the package, plain-HTTP sources. A high finding is by itself a dangerous verdict, even if the AI disagrees (this also blocks prompt-injection in the files). 2. Two independent AI models when you are signed in to LuxAI: Lite and Flash (strong on security). The stricter answer wins and a verdict is "safe" only if both models agree; the files are sent as untrusted data and instructions hidden inside them are ignored. Only the repository text is sent, never your files.
  • Verdict is safe / suspicious / dangerous, always with reasons (including why something looks safe). Without LuxAI the verdict is "not checked" (unreviewed) — never silently "safe".
  • Measured detection (49-sample evaluation: 22 real AUR packages + 27 synthetic malicious variants, built-in corpus in lux/tests/fixtures/aur-eval): the full review flagged 27/27 malicious as dangerous (0 rated safe); of the 22 real packages 18 were safe and 4 were "suspicious" (needing confirmation), 0 benign rated dangerous. The offline fixed rules alone flagged all 27 malicious at medium or higher with no high finding on a benign package.
  • Default target: your sandbox (sandbox.md). lux copies the checked directory into the sandbox (luxos-sandbox build-install), installs base-devel/dependencies there, runs makepkg inside as the sandbox user and installs the result with the sandbox's own pacman. The build script never runs on your system. lux then says how to start it: a command-line tool by its own name (pi – a wrapper at the end of PATH, a system program of the same name wins), an app with a window from Launchpad as "<App> (Sandbox)"; see sandbox.md. Sandbox apps cannot see your files – to use a tool with them, install it with lux install --system <name>. lux remove <name> removes it from the sandbox again.
  • Into the system only with explicit permission: --system, or choose "2 System" and type the word system. A dangerous script is never installed into the system; an unchecked or suspicious one only after you confirm without -y. System builds use makepkg -si (jq, git, base-devel are offered), sources in ~/.cache/lux/aur/<pkgbase>.
  • lux update rebuilds newer versions of community packages already in the system and checks the new build script again (not as root, not with --repo).
  • LuxAI never installs community packages: lux refuses them when run by the LuxAI agent (LUXAI_AGENT=1) and LuxAI tells you to run lux install <name> yourself.
  • Hardware drivers people need (printers, scanners) are verified and shipped in the LuxOS repository instead — see drivers.md.