Posted on Apr 12, 2026 By copyninja under devops

The Shift in Software Consumption

Historically, I have been a "distribution-first" user. Sticking to tools packaged within the Debian archives provides a layer of trust; maintainers validate licenses, audit code, and ensure the entire dependency chain is verified. However, the rapid pace of development in the Generative AI space—specifically with new tools like Gemini-CLI—has made this traditional approach difficult to sustain.

Many modern CLI tools are built within the npm or Python ecosystems. For a distribution packager, these are a nightmare; packaging a single tool often requires packaging a massive, shifting dependency chain. Consequently, I found myself forced to use third-party binaries, bypassing the safety of the Debian archive.

The Supply Chain Risk

Recent supply chain attacks affecting widely used packages like axios and LiteLLM have made it clear: running unvetted binaries on a personal system is a significant risk. These scripts often have full access to your $HOME directory, SSH keys, and the system D-Bus.

After discussing these concerns with a colleague, I was inspired by his approach—using a Flatpak-style sandbox for even basic applications like Google Chrome. I decided to build a generalized version of this using OpenCode and Qwen 3.6 Fast (which was available for free use at the time) to create a robust, transient sandbox utility.

The Solution: safe-run-binary

My script, safe-run-binary, leverages systemd-run to execute binaries within an isolated scope. It implements strict filesystem masking and resource control to ensure that even if a dependency is compromised, the "blast radius" is contained.

Key Technical Features

1. Virtualized Home Directory (tmpfs)
Instead of exposing my real home directory, the script mounts a tmpfs over $HOME. It then selectively creates and bind-mounts only the necessary subdirectories (like .cache or .config) into a virtual structure. This prevents the application from ever "seeing" sensitive files like ~/.ssh or ~/.gnupg.
2. D-Bus Isolation via xdg-dbus-proxy
For GUI applications, providing raw access to the D-Bus is a security hole. The script uses xdg-dbus-proxy to sit between the application and the system bus. By using the --filter and --talk=org.freedesktop.portal.* flags, the app can only communicate with necessary portals (like the file picker) rather than sniffing the entire bus.
3. Linux Namespace Restrictions

The sandbox utilizes several systemd execution properties to harden the process:

  • RestrictNamespaces=yes: For CLI tools, this prevents the app from creating its own nested namespaces.
  • PrivateTmp=yes: Ensures a private /tmp space that isn't shared with the host.
  • NoNewPrivileges=yes: Prevents the binary from gaining elevated permissions through SUID/SGID bits.
4. GPU and Audio Passthrough
The script intelligently detects and binds Wayland, PipeWire, and NVIDIA/DRI device nodes. This allows browsers like Firefox to run with full hardware acceleration and audio support while remaining locked out of the rest of the filesystem.

Usage

To run a CLI tool like Gemini-CLI with access only to a specific directory:

safe-run-binary -b ~/.gemini-config -- npx @google/gemini-cli

For a GUI application like Firefox:

safe-run-binary --gui -b ~/.mozilla -b ~/.cache/mozilla -b ~/Downloads -- firefox

Conclusion

While it is not always possible to escape the need for third-party software, it is possible to control the environment in which it operates. By leveraging native Linux primitives like systemd and namespaces, high-grade isolation is achievable.

PS: If you spot any issues or have suggestions for improving the script, feel free to raise a PR on the repo.