Blast Radius Zero — sandboxed agent runtime

A web agent that plans tasks, fetches pages and runs its own generated code — but every execution happens in a fresh, disposable, network-less Docker sandbox on Vultr. Seven attack probes run in CI and the build fails if any one escapes.

Live demoGitHub0

Overview

Blast Radius Zero is a sandboxed execution agent: you type a task, an LLM plans it into steps, and the code that the model writes is executed — but never on the host.

Every step gets its own container, created for that step and destroyed after it. The container has no network namespace, a read-only root filesystem, a tmpfs workdir, dropped capabilities, a non-root uid, an empty environment, and hard caps on memory, CPU, PIDs, wall clock, output size and artifact disk. Nothing is reused between steps, so a compromised run has nothing to persist into.

The point of the project is not the agent, it is the measurement. A containment lab ships with it: seven attack probes (write to /etc/passwd, fork bomb, runaway loop, phone home, hunt for credentials, memory bomb, disk fill) run against the sandbox, and CI fails the build unless all seven are reported as contained. The result is not a claim in a README, it is an artifact you can reproduce with one command.

A second, independent model call verifies each step's output before the next step is planned, and every run is written to a SQLite ledger with per-step commands and SHA-256 hashes of the artifacts.

Honest boundaries: the control plane itself is root-equivalent because it holds the Docker socket, so a compromise there is a host compromise — this is written up rather than hidden. An experimental WASM backend also exists, but WASI CPython is stdlib-only, so numpy/pandas/matplotlib cannot be imported there; that limitation is documented instead of quietly skipped.

How it uses Vultr

One Vultr Cloud Compute instance (vc2-1c-1gb, Seoul region) runs the whole stack: a single FastAPI process serving the console over HTTPS with SSE streaming, plus the Docker daemon it uses to spawn one throwaway container per step. gVisor (runsc) is used as the container runtime where available, with plain runc as the fallback. Vultr Serverless Inference provides the model behind the OpenAI-compatible /chat/completions endpoint (glm-5.3), used for both planning and per-step verification — so the control plane holds only an inference key and no GPU infrastructure of its own. The sandbox containers inherit no Vultr credentials, no host mounts and no network device: the only thing that crosses into the sandbox is a single program on stdin, and the only things that come back are stdout, stderr, the exit code and up to 25 artifact files. Why Vultr: a $5/month instance is enough to demonstrate the pattern end to end, which makes the security argument reproducible by anyone rather than a whiteboard exercise. TLS is terminated on the instance and valid until 2027-02-01.

No comments yet

Be the first to comment on this project.