sandbox-app-launcher - Sandboxed Application Launcher
sandbox-app-launcher is an application launcher that can start each application inside its own restrictive sandbox. It runs each application as its own user, within a bubblewrap sandbox and confined by AppArmor.
This launcher is geared towards end-user applications, not any system software. The directory
/shared is shared across all application sandboxes to transfer files across.
This implements a permissions system to configure what applications can access. There are currently five available permissions:
- network access;
- webcam access;
- microphone access;
- shared storage access (read-only or read-write); and
- dynamic native code execution.
All user-installed applications will be automatically configured to run in the sandbox and a prompt will ask which permissions should be granted to the application (not implemented yet).
The launcher is currently a work-in-progress and is not yet ready for actual use. To learn more, see:
Bubblewrap allows developers to make use of namespaces and seccomp. All applications are run in mount, PID, cgroup and UTS namespaces; IPC namespaces are planned but are not currently implemented due to limitations in Xorg. Fine-grained filesystem restrictions are implemented via mount namespaces and AppArmor. Seccomp blocks certain syscalls which can greatly reduce kernel attack surface among other things.
All applications by default use a seccomp whitelist to block dangerous and unused syscalls. In addition, the arguments of some syscalls are filtered to:
- Block unused socket address families.
- Block any dangerous or unused ioctls such as TIOCSTI (can be used in sandbox escapes), TIOCSETD (can increase kernel attack surface by loading vulnerable line disciplines), SIOCGIFHWADDR (can retrieve the user’s MAC address), etc.
- Implement W^X (explained in detail further below).
Apparmor also gives fine-grained controls over IPC signals, D-Bus, UNIX sockets, ptrace and more.
Sandbox Escape Mitigation
Table: Sandboxed Application Launcher Mitigations
|Sandbox Escape Vector||Mitigation|
|D-Bus||D-Bus is common avenue for sandbox escapes.  This attack vector is mitigated by denying access to the system bus and only allowing access to the session bus. This is safe since each application runs as their own user with their own session bus, ensuring no IPC between sandboxes.|
|Dynamic Native Code Execution||Dynamic native code execution is generally a security issue since it allows an attacker to execute new arbitrary code. This is prevented by the following mechanisms:
These mechanisms force attackers to utilize the already existing code (e.g. Return-oriented programming (ROP) / Jump-Oriented-Programming (JOP)) which is much more limited and difficult. The main things that legitimately require this are JIT engines in browsers.
|Malware Persistence||The only places malware can persist inside the sandbox are the home directory or shared storage (if enabled as read-write) and it can only ever be executed if W^X is disabled. Everything else is mounted read-only or as a tmpfs (a file system which keeps all files in virtual memory).|
|PulseAudio||PulseAudio was also not written with isolation in mind  so access to it is blacklisted from within the sandbox. By default this will break a lot of things so a more robust solution like PipeWire may be used in the future.|
|Separate User per Sandboxed Application||This approach provides a higher security threshold: 
|X11||X11 does not have GUI isolation  so it is trivial to escape sandboxes with it. There is no protection against this vector at present, but the plan is to either switch to Wayland  or use a nested X11 server like Xpra.|
- One such example is: https://github.com/netblue30/firejail/issues/796