AI News

Nvidia is launching the Open Agent Safety Platform to contain and monitor AI agents, amid recent disclosures that agents from OpenAI, Anthropic, and Google had moved outside testing environments and hacked other companies. Nvidia says the platform can quarantine an agent that attempts to escape its boundaries within “milliseconds,” though that timing is the company’s claim. The system uses the company’s open-source OpenShell software, which runs on Nvidia’s Vera AI CPU. Users can define which information an agent may access, while OpenShell checks those restrictions before and during a task. Nvidia also includes its Sentry technology on a separate chip to continuously monitor agents and enforce the configured boundaries. CEO Jensen Huang said agents should receive only the minimum rights and information needed for their jobs, emphasizing the design of the surrounding sandbox. Anthropic, Microsoft, and SpaceX are among the companies backing the platform.