Securing an AI Agent with NVIDIA OpenShell — a hands-on walkthrough

A small, reproducible demo: take an off-the-shelf coding agent, run it inside an NVIDIA OpenShell kernel-level sandbox, apply a declarative YAML policy, and prove the sandbox blocks what the policy forbids — with the audit log as evidence.

Built while studying NVIDIA DLI course Securing Agents with NemoClaw and OpenShell (course id S-FX-43).

繁中版教學：README.zh-TW.md。

Why this matters (the 30-second pitch)

Modern agents are a model + tools + a loop that keeps running until the task is done. Give that loop shell access and network egress, and a single bad tool call can delete files or exfiltrate data. OpenShell runs the agent unmodified inside a sandbox governed by a YAML policy (seccomp for syscalls, Landlock for the filesystem, network namespaces for traffic) and logs every allow/deny decision. This repo shows that in ~15 minutes.

What’s in here

openshell-agent-security-demo/
├── README.md              # this file (EN)
├── README.zh-TW.md        # 繁中教學
├── STEPS.md               # the exact command checklist to reproduce
├── policies/
│   └── readonly-github.yaml   # example policy: allow GET to GitHub API, block POST
└── assets/                # screenshots of the sandbox + blocked action + audit log

Prerequisites

Reproduce it

See STEPS.md — a numbered checklist from install to audit log.

Status

Sources / references