Nvidia Launches Silicon-Based Governance Platform for AI Agents

Nvidia has launched the Open Agent Safety Platform, a governance system designed to monitor and secure autonomous AI agents through a combination of software and hardware-level enforcement. The platform aims to provide full-stack control across the compute and robotics systems that power agentic AI.

The system integrates OpenShell software with a hardware component known as NVIDIA Sentry. Sentry operates as an out-of-band watchdog running on NVIDIA BlueField-4 data processing units (DPUs). This setup is designed to create a secure runtime environment where AI agents operate within sandboxed environments featuring kernel-level isolation.

Hardware-Level Enforcement via NVIDIA Sentry

A central feature of the platform is its ability to provide in-silicon security enforcement. Nvidia stated that if an AI agent attempts to breach its defined software boundaries, Sentry can quarantine and stop the agent within milliseconds. The company argues that by using OpenShell, it can provide a secure runtime boundary that traces all actions and enforces policy as agents run on Nvidia Vera CPUs.

Frank Dickson, a principal analyst at Dickson Research, suggested that the hardware-centric approach addresses a fundamental flaw in previous AI security attempts. Historically, many efforts tried to secure agents by implementing guardrails within the model or the prompt itself. “That approach was always going to lose,” Dickson said. “Nvidia finally puts enforcement where it belongs: outside the agent, in the kernel, in a proxy that inspects every outbound request and in silicon the agent cannot see or touch.”

The platform is being rolled out alongside an extensive list of partners, including Anthropic, Cisco, CrowdStrike, Microsoft, Salesforce, and ServiceNow. However, analysts noted that major industry players, including OpenAI, Amazon, and Google, are notably absent from the current list of supporters.

Limitations in Coverage and Vendor Lock-in

Despite the technical advancements, cybersecurity experts have warned that the platform does not address the full spectrum of enterprise risks. A primary concern is the issue of “shadow AI”—the proliferation of unapproved agents within an organisation.

Brian Levine, a partner at the consulting firm Control Risks, noted that these controls are effective only for agents deployed on infrastructure that an enterprise directly manages. The platform provides no visibility or control over agents spun up on third-party SaaS platforms or those embedded within a vendor’s existing products. “You can’t hold an agent to a policy if you don’t know it exists,” Levine said, highlighting that discovery and inventory remain significant governance hurdles.

Brent Ellis, VP of AI Infrastructure at IDC, estimated that the platform likely addresses less than 25% of the total cybersecurity challenges associated with enterprise agentic AI. He also cautioned that the offering introduces a risk of vendor lock-in, as the security benefits are heavily tied to the adoption of Nvidia’s specific silicon architecture.

Furthermore, security professionals warned that hardware-enforced boundaries are not a total solution. Justin Greis, CEO of Acceligence, pointed out that agents may still exploit unintended pathways or rely on misconfigured permissions and excessive authority to cause damage, even without breaking through the primary sandbox wall.