NVIDIA/OpenShell
A runtime that runs AI agents inside an isolated sandbox. You declare which files, system calls and network hosts the agent may touch. API keys go only into requests bound for approved hosts, and any change to the policy waits for a human to review it.
Not tested yetTested by others
How we know
Someone outside the project tested it. Their result is below, with a link to the original post. We have not tested it ourselves.
Good for
Teams running autonomous agents on their own machines or in Kubernetes who need hard limits on what an agent can reach.
Not for
Solo use with a few agents on a laptop, or anyone without Docker, Podman or a VM. It is still 0.1.x, so it is not yet proven for production.
Sources
malicious setup script leaked a canary secret in 10 of 10 runs without OpenShell and 0 of 10 under the default policy; automatic approval granted new public hosts with no human in 12 of 12 trials
Sorami · · how: v0.1.2 on one Apple Silicon Mac with the microVM driver only, 35 test IDs, 123 trials, test plan committed before the scored runs; the agent in the leak test was a small local model (qwen3:8b)
https://sorami.com.au/research/nvidia-openshell-agent-sandbox-test/
Repo
- Stars
- 12,648
- License
- Apache-2.0Use it, change it, sell it, closed-source included. Keep the author's copyright notice.
- Maturity
- active
- Last push
- 2026-10-01
- Last checked
- 2026-11-14