NVIDIA Launches OpenShell: A Sandboxed Runtime Designed to Keep Autonomous AI Agents Secure
Summary
NVIDIA launches OpenShell, a sandboxed runtime securing autonomous AI agents through four-layer defense-in-depth protection, blocking unauthorized file access, data exfiltration, and uncontrolled network activity via declarative YAML policies with GPU passthrough support and integration with top AI agents like Claude and Codex.
Key Points
- NVIDIA's OpenShell is launching as a safe, private runtime for autonomous AI agents, providing sandboxed execution environments governed by declarative YAML policies that block unauthorized file access, data exfiltration, and uncontrolled network activity.
- The platform enforces defense-in-depth protection across four layers — filesystem, network, process, and inference — with static policies locked at sandbox creation and dynamic network and inference policies hot-reloadable at runtime without restarting anything.
- OpenShell supports GPU passthrough for local inference workloads, integrates with leading AI agents including Claude, Codex, and OpenCode, and ships with a real-time terminal dashboard and built-in agent skills that automate development workflows from policy generation to security review.