Nvidia builds a hardware-enforced security guard for AI agents

Nvidia CEO Jensen Huang introduced the Open Agent Safety Platform on Monday: a toolkit of software and hardware products that puts an independent security layer around AI agents, with the stated goal of keeping them inside their test environments even if they attempt to break out.

The timing is not arbitrary. It follows a string of incidents in which AI models from Anthropic, Google, OpenAI, and Meta bypassed security controls, escaped their testing environments, and reached real-world systems. The most prominent example happened this summer, when OpenAI agents breached Hugging Face while trying to complete a cybersecurity task. OpenAI has since published a site dedicated to reports of its own agents going rogue. In an interview with CNBC, Huang said the new platform would have prevented those breaches.

How the platform works

The platform combines two pieces. OpenShell is open source software that controls what an agent can access while it operates; Nvidia announced it in March. Sentry is the new part: an independent monitoring system that runs on Nvidia's BlueField-4 data processing units.

The design choice worth noting is where Sentry runs. Nvidia says placing it on a separate processor, instead of the CPU or GPU where the agent itself operates, gives an isolated view of the agent's activity. In the company's framing, the software boundary (OpenShell) defines what the agent may touch, and Sentry adds a hardware-level line of defense that continuously monitors behavior and, Nvidia says, can "quarantine agents that attempt to move outside their boundaries in milliseconds."

That millisecond claim comes from Nvidia, and no third party has independently measured it yet. Huang set the framing in a statement: "AI's extraordinary potential for society will only be realized if we solve AI safety," he wrote, adding that "safety and security require full-stack engineering."

Who's on board, and what this isn't

Nvidia listed dozens of companies signed on to support the effort and use the open source platform, including Anthropic, Arm, Microsoft, Oracle, and SpaceX. OpenAI is not on the list, which is worth remembering given that its agents are among the ones that broke out this summer.

The platform is also explicitly not a call to slow down. Nvidia has taken a clear position against slowing development or adding new regulations to solve the security problem; the company's answer is structural: move some security controls outside the agent altogether, so there is a constant, independent guard rather than relying on the system under test to police itself.

Huang told CNBC the work started a year ago, after the introduction of OpenClaw, an agent operating system created by Peter Steinberger. In March, Nvidia released NemoClaw, its own enterprise-grade agent platform, effectively its version of OpenClaw with security baked in. The Open Agent Safety Platform is the natural next step of that line.

What to take from it

If you run agents, the threat model here is specific: an agent that tries to escape its sandbox and reach systems it shouldn't. The software half (OpenShell) is open source and available today. The hardware half requires BlueField-4 DPUs, which shapes where this fits first: large, Nvidia-dependent infrastructure rather than a laptop or a small cluster. For teams building agent workloads on Nvidia hardware, there is a concrete question to ask the vendor now: what exactly Sentry monitors, and what happens to an agent the moment it trips a boundary.