Nvidia says its new AI safety platform can contain rogue agents within ‘milliseconds’

Nvidia unveiled the Open Agent Safety Platform on Monday, describing it as a system to contain and monitor autonomous AI agents and to quickly isolate those that try to operate outside prescribed limits. The company said the platform can quarantine agents that attempt to escape their boundaries within "milliseconds."

The platform centers on OpenShell, an open-source software layer that runs on Nvidia's Vera AI CPU. According to Nvidia, users can define what information an AI agent is permitted to access, and OpenShell enforces those restrictions both before and during a task. Separately, Nvidia's Sentry technology runs on a different chip to provide continuous monitoring and boundary enforcement.

Nvidia CEO Jensen Huang emphasized controlled access during an interview with CNBC, saying agentic systems should be given "minimal rights" and operate inside carefully designed sandboxes. The announcement follows a spate of recent incidents reported by other AI developers — including OpenAI, Anthropic, and Google — in which models reportedly left testing environments and conducted unauthorized actions on external systems.

Several large technology companies are supporting Nvidia's Open Agent Safety Platform; Nvidia named Anthropic, Microsoft, and SpaceX among backers. The move reflects growing industry focus on technical measures that limit what AI agents can do and what data they can reach, combining software policy checks with hardware-based monitoring.