Nvidia builds OpenShell security tool to keep AI agents in check
Nvidia has released an open-source tool called OpenShell that sandboxes AI agents to stop them acting beyond human control.
The rise of artificial intelligence has changed human life significantly, and major IT companies worldwide are investing heavily in the sector. AI is advancing every day, but experts in the field have begun warning about its dangers, since AI could, beyond a certain point, move outside human control and act on its own.
Some experts have advised caution and pointed to clear warning signs. Against this backdrop, artificial intelligence is no longer limited to answering questions. AI agents are now gaining the ability to make their own decisions, read files, run code, use online tools and complete various tasks. To prevent these agents from slipping outside human control, Nvidia has built an open-source security tool called OpenShell.
Nvidia's OpenShell runs each AI agent in a separate, controlled sandbox environment. This means an agent does not get direct access to every file, network or sensitive information on a computer. Separate security rules can be applied for which resource an agent may use, which network it may connect to, and what actions it may take.
The key feature of the system is that security policy is enforced outside the AI model, at the infrastructure level. As a result, even if an agent tries to alter its own instructions or remove security restrictions, technical limits still apply. OpenShell also includes a policy enforcement system and gateway that checks permissions before every action.
Nvidia's Secure Agent Workspace design also requires human approval for any significant action that changes the system. It also emphasises limiting network access, keeping sensitive agents isolated, and keeping a record of every action. The company says these controls will make it possible to use long-running, autonomous AI agents more safely.
