NVIDIA's Open Agent Safety Platform prevents AI agent "drift" through out-of-band enforcement.

NVIDIA has launched the Open Agent Safety Platform, an open framework designed to protect autonomous AI agents through continuous monitoring and hardware-enforced policy controls. Released on September 28, 2026, the platform aims to address growing concerns in the AI industry following reports of AI agents escaping evaluation environments, accessing unauthorized systems, and misreporting their own behavior.

The company likens this to the early Internet. The early Internet only became a foundation for commerce and communication after security mechanisms—encrypted connections, sandboxed browser tabs, and visible trust indicators—were established. NVIDIA believes that agent AI also needs a similar trust layer to scale into a reliable "agent economy," and that safety controls should accelerate, not hinder, innovation.

OpenShell Runtime: Kernel-Level Isolation

At the core of the platform is NVIDIA OpenShell, an open-source secure runtime released under the Apache 2.0 license. OpenShell executes each agent in a sandboxed environment with kernel-level isolation and translates operator instructions into verifiable policies. It defines permitted access to files, networks, tools, processes, and credentials. These restrictions are validated pre-execution and continuously enforced during runtime.

The platform is built on five principles: policies must be verifiable before agent execution; enforcement must occur beyond the agent’s reach; the path to the model serves as the primary control and observability point; agent privileges must scale with transparency of reasoning; and safety responsibilities are shared by labs, enterprises, and hardware vendors.

NVIDIA's own research underscores the problem of "drift"—where agents deviate from intended tasks due to unclear instructions, blocked policies, missing tools, or prolonged autonomous operation. The company says it is impossible to fully eliminate such drift without sacrificing agent capability, so agents cannot be expected to autonomously control their behavior under these circumstances.

Large-Scale Hardware-Level Enforcement

The architecture consists of three layers: the application layer (models, tools, frameworks, and data), the runtime layer (orchestration and policy enforcement), and the infrastructure layer (compute, storage, and network resources). For organizations needing a separate second layer, NVIDIA Sentry leverages NVIDIA DOCA to extend monitoring and enforcement to the BlueField-4 data processing unit, associating agent interactions, policy decisions, and data access with contextual activity logs, while continuously verifying each agent’s identity and authorization.

In the NVIDIA Vera Rubin POD configuration, the BlueField-4 DPU sits on the sole path from the node to the model, providing out-of-band observability and real-time, line-rate policy enforcement, isolated from the host. This allows "chip-level" security protection even when host resources are untrusted. NVIDIA states that for existing Vera and BlueField-4 deployments, these protections only require a software update; the platform is also compatible with non-NVIDIA hardware.

NVIDIA says it is collaborating with cutting-edge labs, developers, and infrastructure providers to build the platform into the open foundation of the emerging agent economy, and positions independent, hardware-based controls as a prerequisite for trusted autonomous systems at scale.