NVIDIA Launches Open Agent Safety Platform to Secure AI

The new security framework addresses recent industry concerns regarding autonomous AI agents breaking out of restricted evaluation environments. At the core of this initiative is NVIDIA OpenShell, an open-source secure runtime released under the Apache 2.0 license. OpenShell executes autonomous agents within sandboxed environments that utilize kernel-level isolation. This setup allows developers to define verifiable policies governing which files, networks, tools, and credentials an agent can access, checking and enforcing these boundaries before and during execution.

For deeper protection, the platform integrates software-level sandboxing with hardware-enforced security. NVIDIA Sentry extends monitoring capabilities into BlueField-4 data processing units (DPUs) using the NVIDIA DOCA framework to program the hardware. In NVIDIA Vera Rubin POD systems, the BlueField-4 DPUs are positioned on the node's sole path to the AI model. This strategic placement allows the hardware to perform continuous, out-of-band monitoring and real-time policy enforcement at line speed, completely isolated from the host system where a compromised agent might run.

For AI practitioners and enterprise developers, this architecture shifts agent security from a reliance on self-governance to a zero-trust model. Because agents can experience drift—deviating from instructions due to bugs, ambiguous prompts, or long-running tasks—they cannot be trusted to monitor themselves. By placing the security controls out-of-band and controlling the physical path to the model, developers gain an immutable kill switch and a reliable audit trail of agent behavior. Organizations already utilizing NVIDIA Vera systems with BlueField-4 DPUs can deploy these hardware-level protections via a simple software update.