AI Card Bao AICARDBAO.COM

Industry News

NVIDIA has created a “prison” for AI agents – called OpenShell, and its current version number is still 0.0.

Running a programming agent directly on one’s own machine always carries some sense of concern: after all, it needs to read files, install dependencies, call APIs, and use keys. If too few permissions are granted, the agent won’t be able to function properly; whereas if too many permissions are given, it might one day be used for malicious purposes, with potentially serious consequences.

NVIDIA’s open-source project OpenShell is precisely designed for this purpose. Some people have described it as a “prison” created specifically for AI agents.

The mechanism by which it manages proxies consists of two layers. The first layer involves the enforcement of policies at the kernel level during runtime; all file accesses, system calls, and network connections must go through this process. The second layer is formal verification: before making any changes to the policy, it is necessary to assess whether such modifications will introduce new security risks or grant unnecessary privileges. This step helps prevent the inadvertent creation of high-risk vulnerabilities in an effort to simplify procedures.

The second aspect is what distinguishes it from ordinary sandboxes. In most cases, the permissions granted to users become increasingly extensive over time—each time an error occurs, another permission is added. After several months, no one can anymore clearly remember what specific privileges have been granted.

Permissions can be set with great precision. For example, it is possible to allow access to the GitHub API while prohibiting any posting operations.

The support is also quite extensive – tools such as Claude Code, Codex, OpenCode, and GitHub Copilot CLI can all be integrated, in addition to solutions like Ollama, OpenClaw, and Hermes Agent. Last October, NVIDIA further developed an enterprise reference architecture based on this foundation, which includes agents from eight major model manufacturers, including Anthropic, Cohere, Google, Meta, and OpenAI.

There is indeed cold water to be poured. Project version numbers start at 0.0, and there have been very few publicly available production implementations. For kernel-level components, the cost of troubleshooting issues can be quite high; individual developers often find it too cumbersome to go through the process of installing such systems just to use a proxy.

Formal verification can only tell you whether your permissions have been expanded or not; it cannot detect whether a proxy is performing inappropriate actions within the authorized scope. If you have granted the permission to delete a database, then that proxy will still be able to do so.

Related