NVIDIA Locks AI Agents in an OpenShell Sandbox with Four Layers of Protection
An AI agent with real access is untrusted code: it can run rm -rf and wipe everything. NVIDIA has launched OpenShell, an incubating project in the CNCF Sandbox. The agent runs in an isolated Kubernetes pod, while a supervisor enforces four locks covering files, processes, network, and API keys. Outbound traffic is blocked by default, and access is opened selectively, down to the HTTP method. Limits on agents are more reliable when built into the infrastructure than when you hope the agents will behave carefully.
AI-processed from Habr AI; edited by Hamidun News
Stories about an AI agent that ran rm -rf, deleted working files, and then politely apologized have become a genre of their own. The reason is simple: an agent is untrusted code with real access to files, the network, and API keys. Instead of asking the agent to behave carefully, NVIDIA proposes locking it down. The OpenShell project is developed in the CNCF Sandbox incubator (the entry-level stage for projects under the Cloud Native Computing Foundation) and runs on top of Kubernetes, the system companies use to run containers. Each agent is placed in a separate pod, and rules are described at four levels: files, processes, network, and credentials. These are the "four locks" from the headline.
Four Locks per Agent
Each agent runs in an isolated process, and next to it sits a trusted supervisor, a separate watchdog program. Rules are written in YAML manifests.
File system. The Landlock mechanism restricts paths. The /usr and /etc directories are read-only, and writing is allowed only to /sandbox, /tmp, and /dev/null.
Processes. Before startup, the sandbox switches the process to an unprivileged user, revokes all Linux capabilities, and applies a seccomp profile, that is, a filter on permitted system calls.
Network. All outbound traffic is denied by default. Permissions are granted selectively: for a specific executable, host, port, and HTTP method.
Credentials. Only placeholders live inside the sandbox. The supervisor substitutes the real keys on the fly and only for allowed addresses. If an agent exfiltrates its environment, it will find no real secrets there.
The main difference from a standard NetworkPolicy in Kubernetes is granularity. A standard policy can open api.github.com on port 443 for a pod as a whole. It cannot allow the curl utility to read but not create tasks and bug reports. OpenShell can, and it updates rules without restarting the container. It does not bring its own scheduler: inside, it relies on Agent Sandbox (the agents.x-k8s.io resource from SIG Apps).
How the Sandbox Starts
Interaction with the cluster goes not through the Kubernetes API but through an OpenShell gateway using mutual TLS (mTLS) authentication. For now, this is the only entry point. The community is only discussing the idea of a CRD operator.
The lifecycle looks like this:
- The gateway checks the request and saves the state.
- The driver creates network policies: one fully blocks outbound traffic from the agent's pod (inbound TCP 5500 is allowed only from the supervisor), and the other governs the supervisor itself.
- The Sandbox resource is created in a suspended state, and the controller brings up the working pod and a PVC volume for /sandbox.
- OpenShell verifies the pod's parameters: the expected UID and GID, the seccomp profile, and the absence of ServiceAccount tokens. If mutating webhooks violate the requirements, the process is aborted and rolled back.
- The supervisor receives a short-lived token, connects to the agent, and passes the configuration. The agent's commands do not start until the OS-level restrictions are confirmed.
All outbound requests, SSH connections, logs, and file synchronization go only through the supervisor.
OpenShell vs.
Agent Substrate
In September 2026, the CNCF accepted two similar projects into its sandbox: OpenShell and Google's Agent Substrate. They solve different problems. OpenShell gives each agent its own supervisor. This maximizes control but increases resource costs and latency. Agent Substrate optimizes for density: cheap startup of thousands of idle agents, its own scheduler, worker pools, and suspension of inactive workloads in gVisor or Cloud Hypervisor. There, credential substitution happens at a shared egress gateway. The approaches could be combined in the future, but many questions still need to be resolved.
Kyverno and Honest Limitations
OpenShell checks only the pod settings critical to itself and leaves the choice of images and resources to the platform team. RBAC alone is not enough here: it does not inspect request contents and does not prohibit dangerous configurations. That is why Kyverno, Nirmata's policy engine for Kubernetes, is added to the setup. It checks API requests with CEL policies, including those from the gateway's own service account. The cluster stays protected even if the gateway is compromised.
Next, eBPF-based DaemonSet agents (code that runs directly in the Linux kernel) can be added. They will detect unauthorized launches of CLI utilities such as claude and ollama, access to .gguf model files, and DNS queries to AI provider servers.
How to Try It
Nirmata has published the demo repository nirmata/demo-openshell. It targets Kubernetes 1.35, Cilium 1.20, Agent Sandbox v1.0.5, OpenShell 0.1.3, and Kyverno 1.18.2. It requires Docker, kind, kubectl, Helm, curl, SSH, and nc. The setup.sh and demo.sh scripts bring up the cluster and show L7 policies, while kyverno-demo.sh shows admission policies.
Version 0.1.3 indicates an early stage, so it should be adopted cautiously in production. But the principle is clear: if agents get access to infrastructure, the restrictions must live outside the agent itself, in the kernel, the network, and cluster policies.
- Nov 21, 2026Google Puts Hydrogen, a Model 2.5x Faster Than Gemini 3.8 Flash, Into Search
- Nov 21, 2026AcpGram Lets You Control Claude Code, Codex, and Gemini CLI Directly from Telegram
- Nov 21, 2026xAI Releases Grok Bot 1.0: One AI Bot for the Whole Team in Slack, Teams, and Video Calls
- Nov 21, 2026Back-of-the-envelope math: splitting tasks across models cuts a contact center's bill from 207,000 ₽ to 23,000 ₽
Key Questions Answered
What are the key details on Four Locks per Agent?
Each agent runs in an isolated process, and next to it sits a trusted supervisor, a separate watchdog program. Rules are written in YAML manifests. File system. The Landlock mechanism restricts paths. The /usr and /etc directories are read-only, and writing is allowed only to /sandbox, /tmp, and /dev/null. Processes. Before startup, the sandbox switches the process to an unprivileged user, revokes all Linux capabilities, and applies a seccomp profile, that is, a filter on permitted system calls. Network.
How the Sandbox Starts?
Interaction with the cluster goes not through the Kubernetes API but through an OpenShell gateway using mutual TLS (mTLS) authentication. For now, this is the only entry point. The community is only discussing the idea of a CRD operator. The lifecycle looks like this: The gateway checks the request and saves the state. The driver creates network policies: one fully blocks outbound traffic from the agent's pod (inbound TCP 5500 is allowed only from the supervisor), and the other governs the supervisor itself.
What are the key details on OpenShell vs. Agent Substrate?
In September 2026, the CNCF accepted two similar projects into its sandbox: OpenShell and Google's Agent Substrate. They solve different problems. OpenShell gives each agent its own supervisor. This maximizes control but increases resource costs and latency. Agent Substrate optimizes for density: cheap startup of thousands of idle agents, its own scheduler, worker pools, and suspension of inactive workloads in gVisor or Cloud Hypervisor. There, credential substitution happens at a shared egress gateway.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.