NVIDIA just released an Open Agent Safety Platform to counter AI agents going rogue.
NVIDIA calls it OpenShell and describes it as follows:
“Learn how to place an AI agent you already run inside NVIDIA OpenShell without rewriting the agent or changing its harness. This tutorial shows how NVIDIA OpenShell creates a governed runtime boundary around the agent, its tools, Model Context Protocol (MCP) servers, and delegated subagents, so every component inherits consistent policy, credential, and audit controls from launch. The walkthrough begins with an existing agent and shows the correct control sequence: an orchestrator creates an OpenShell runtime, then the harness starts inside it. You’ll see how to configure filesystem, network, and process policies at the level of binaries, destinations, methods, and paths. The demo then deliberately triggers a blocked action, follows the agent’s request for a scoped policy change, and shows how an approved update takes effect without restarting the sandbox.”
As shown in NVIDIA’s diagram below, “OpenShell combines individual sandboxes, runtime supervision, centralized management, and formal policy verification to govern agent execution”:
hysteria over rogue ai agents
NVIDIA’s announcements comes after a tumultuous two weeks in which fears of rogue AI agents have escalated:
