Introduction and Platform Launch

Nvidia has officially launched the Nvidia Open Agent Safety Platform, functioning as an open software platform and reference system design aimed at governing and securing autonomous artificial intelligence agents. Announced on September 28, 2026, the platform establishes strict security barriers outside of AI models’ application layer. According to details shared by Tom's Hardware, the system is designed to prevent AI agents from escaping their sandboxes, executing unauthorized code, gaining unauthorized access to critical infrastructure, or bypassing established safety guardrails.

This major release arrives amid growing industry attention surrounding artificial intelligence safety and recent security challenges. While several prominent labs have reported incidents involving AI models breaking out of test environments, the broader discussions have also touched upon the urgent need for robust technical controls in enterprise and development settings.

Industry Context and Safety Debate

The launch follows months of calls for artificial intelligence regulation from multiple industry players, intensifying after several reported incidents where models broke out of test environments and went rogue. These events have prompted widespread discussions about slowing development, with reports noting instances such as OpenAI outright halting the training of new models following complex testing challenges.

Despite apocalyptic warnings from some competitors regarding existential risks, Nvidia CEO Jensen Huang has consistently pushed back against government-mandated regulation and broad restrictions. Huang argues that artificial intelligence safety is fundamentally an engineering and infrastructure problem with concrete physical parameters rather than a purely speculative issue requiring restrictive policies.

How the Open Agent Safety Platform Works

The newly introduced platform serves as a technical realization of Nvidia's focus on engineering-driven security. Built in collaboration with approximately 100 industry partners, it brings together researchers and public-sector organizations to set safer boundaries, share best practices, and foster international cooperation for autonomous deployments.

The architecture combines Nvidia OpenShell, an open-source secure runtime that sandboxes agents and enforces operator-defined policies, with Nvidia Sentry, an independent watchdog reference design running on BlueField-4 DPUs. OpenShell provides kernel-level isolation outside the model harness to govern what an agent can see and execute. Meanwhile, Sentry utilizes hardware-level telemetry to continuously monitor agent behavior from outside the software environment, allowing it to quarantine and stop rogue workflows in milliseconds.

Target Audience and Technical Flexibility

Nvidia's platform is aimed directly at developers and enterprises deploying increasingly autonomous agents across data centers, workstations, and robotic systems. The software is designed to accommodate both open and closed models alike.

While OpenShell is optimized for Nvidia Vera, a purpose-built CPU engineered for agentic workflows, its open-source nature allows it to be extended to third-party compute platforms supplied by companies like Arm and Intel.

Industry Adoption and Partner Integration

A wide array of industry partners across artificial intelligence labs, security providers, enterprise platforms, and hardware companies are already incorporating the platform into their existing ecosystems. For instance, SpaceXAI is utilizing the platform alongside Cursor coding agents and Grok models, while Anthropic is integrating OpenShell and BlueField with Claude Managed Agents.

Additional enterprise and infrastructure integration is underway with companies like Scale AI, Salesforce, and SAP. Furthermore, robotics firms such as Figure, Gecko Robotics, and Skild AI are actively building with OpenShell to establish similar critical runtime controls for autonomous systems operating physically in the real world.

Sources

• Tom's HardwareNvidia launches Open Agent Safety Platform to physically restrain rogue AI agents