Okay, supply chain attacks are spreading as planned (trivy, litellm, axios [...]). Producing code has never been that affordable, and projects are increasingly filled with AI-generated artifacts. It might be the best time to inject malicious code.

grumpycat (leslie-alexandre d.)

@foobar.eagleusb.com

3mo

The consequences of axios being compromised are massive. Brace brace. github.com/axios/axios/...

axios@1.14.1 and axios@0.30.4 are compromised · Issue #10604 · axios/axios

more details: https://www.stepsecurity.io/blog/axios-compromised-on-npm-malicious-versions-drop-remote-access-trojan Most likely, a maintainer's GitHub and npm accounts are compromised as these iss...

github.com

1

While many of us are executing privileged processes everywhere, hoping for the best, others remain skeptical about adopting agent(s) in critical production paths without good measure.

Your systems contain secrets, documents, and other sensitive data you don't want leaked or altered.

Sandboxing Tooling

Not an exhaustive list, just mine.

1. 1.

   https://github.com/anthropic-experimental/sandbox-runtime

   • process isolation on multi-platforms; works not only with agent but any process

   • support custom network proxy implementation

   • native process sandboxing with seatbelt (macOS) & bubblewrap (Linux)

   • as (almost) easy as npm install -g @anthropic-ai/sandbox-runtime (you should npm config set prefix "~/.local")

2. 2.

   https://github.com/always-further/nono

   • process isolation on multi-platforms; works not only with agent but any process

   • native process sandboxing with seatbelt (macOS) & bubblewrap (Linux)

   • already designed with agents in mind nono run --profile opencode -- opencode

   • better proxy and networking design

   • Rust

To go further, other projects are aiming isolation in other contexts or doing the same.

• https://github.com/alibaba/OpenSandbox

  • more an SDK to build on top

  • Kubernetes / containers primitive for sandboxing

• https://github.com/kubernetes-sigs/agent-sandbox

  • CRD dedicated to Kubernetes, introduce unique stateful pod declaration instead of current SatefulSets

• https://github.com/Use-Tusk/fence - https://github.com/GreyhavenHQ/greywall

  • native process isolation with bubblewrap

  • both have the same design

• https://github.com/sandboxec/sandboxec

  • mostly one man project

• https://github.com/NVIDIA/OpenShell

  • isolation based on containers

First things first, I shortlisted srt and nono. I choose the former for its Anthropic support and the latter for its native Rust implementation plus advanced features.

srt

• isolation based on user namespace that I already have available

• non-invasive in my local workflow

• (evil) big corp supporting the project

• ease of migration if needed

Configuration

~/.srt-settings.json

Deny by default, allow what necessary in the current directory. Will be enhanced.

{
  "network": {
    "allowedDomains": [
        "github.com",
        "*.github.com",
        "lfs.github.com",
        "api.github.com",
        "npmjs.org",
        "*.npmjs.org",
        "*.golang.org",
        "*.z.ai"
    ],
    "deniedDomains": [
        "packages.npm.org",
        "sfrclak.com"
    ]
  },
  "filesystem": {
    "denyRead": ["~"],
                "allowRead": [
                        "~/.config/opencode/",
                        "~/.local/share/opencode/",
                        "~/.local/bin/",
                        "~/.flox/run/x86_64-linux.default.run/bin/",
                        "~/.local/lib/node_modules/",
                        "~/go/",
                        "/tmp",
                        "."
                ],
    "allowWrite": [
                        "~/.config/opencode/",
                        "~/.local/share/opencode/",
                        "~/go/",
                        "/tmp",
            "."
    ],
    "denyWrite": [".env"]
  },
  "enableWeakerNestedSandbox": false,
  "enableWeakerNetworkIsolation": false
}

Just sandbox by default, every time.

~/.bash_aliases

alias opencode='srt -- opencode'

Verify what your agent attempts to do, easily, using strace. I have focused on ~ file operations, where I utilize custom binaries (nix with flox.dev) along with Node.js, Go, and other libraries. strace is very practical for designing your profile (~/.srt-settings.json)

strace -e file -- opencode --version 2>&1 | grep '/home/'

execve("/home/p00/.local/bin/opencode", ["opencode", "--version"], 0x7ffd7c23d140 /* 156 vars */) = 0
openat(AT_FDCWD, "/home/p00/.local/lib/ollama/glibc-hwcaps/x86-64-v3/libselinux.so.1", O_RDONLY|O_CLOEXEC) = -1 ENOENT (Aucun fichier ou dossier de ce nom)
newfstatat(AT_FDCWD, "/home/p00/.local/lib/ollama/glibc-hwcaps/x86-64-v3/", 0x7fff890b2820, 0) = -1 ENOENT (Aucun fichier ou dossier de ce nom)
openat(AT_FDCWD, "/home/p00/.local/lib/ollama/glibc-hwcaps/x86-64-v2/libselinux.so.1", O_RDONLY|O_CLOEXEC) = -1 ENOENT (Aucun fichier ou dossier de ce nom)

See you in a bit ✌︎㋡