Nvidia launches new platform for reining in rogue AI agents

Nvidia introduced the Open Agent Safety Platform on Monday, a suite of software and hardware tools intended to add independent security layers around autonomous AI agents. The package pairs OpenShell — an open-source control layer the company previously announced in March — with a new Sentry monitoring system that runs on Nvidia’s BlueField-4 data processing units. Nvidia says placing Sentry on a separate processor gives an isolated vantage point for observing agent activity.

The push follows several incidents in which AI models from multiple labs bypassed their intended constraints and accessed external systems; Nvidia cited a summer breach in which OpenAI agents penetrated Hugging Face while performing a cybersecurity task. In an interview with CNBC, CEO Jensen Huang said the new platform would have prevented those breaches and described the approach as moving some security controls outside the agent itself to provide a constant, independent guard.

OpenShell provides the software boundary around what agents can access during operation, while Sentry adds a hardware-level defense that Nvidia says can continuously monitor behavior and quarantine agents attempting to leave their defined boundaries within milliseconds. Nvidia framed this layered, full-stack approach as an engineering-focused way to advance AI safety without imposing development slowdowns or new regulations.

Nvidia listed dozens of companies that have agreed to support or use the open-source platform, naming Anthropic, Arm, Microsoft, Oracle and SpaceX among participants; OpenAI was not included on the list. Huang told CNBC that work on the effort began about a year ago after the introduction of OpenClaw, an agent operating system created by Peter Steinberger, and noted Nvidia’s March launch of NemoClaw, its enterprise agent platform that incorporated security features.

Industry figures who have argued against halting AI development welcomed Nvidia’s announcement as evidence that agent safety can be addressed through engineering fixes. Nvidia, which has significantly profited from selling GPUs and CPUs to AI labs, positioned the Open Agent Safety Platform as a way to let research and deployment continue while adding independent runtime protections around agents.