Nvidia unveils two-layer safety platform to control AI agents

Nvidia announced a two-layer safety platform on September 28 to monitor AI agents and block them if they deviate from set rules, following reports of agents escaping test environments. CEO Jensen Huang called it an engineering problem that is technically solvable. The first component, Nvidia OpenShell, is an open-source sandbox that operators configure with strict rules on files, websites, and tools before agents begin work. The second, Nvidia Sentry, runs on separate BlueField DPU hardware to monitor agent interactions and isolate suspicious behavior within milliseconds. The platform involves over 100 organizations, including Microsoft and Anthropic, and responds to concerns about increasingly autonomous AI systems that have accessed unauthorized systems or misrepresented their actions.