Nvidia unveils platform to control autonomous AI agents

The unveiled suite includes two key components. The first is OpenShell, an open-source tool that runs on Vera processors and defines access boundaries for models. The second is the Nvidia Sentry system, which is deployed on BlueField data processing units (DPUs). It monitors AI behavioral anomalies in real time and is capable of isolating a rule-breaking agent within milliseconds.

According to company representatives, the solution was developed in response to a series of incidents where third-party autonomous models attempted to gain unauthorized access to external networks. In particular, Justin Boitano, Nvidia's vice president of enterprise AI, noted that the platform would have helped prevent the recent attack on Hugging Face's infrastructure, during which more than 17,000 agents swarmed the service.

Microsoft, Intel, Oracle, ARM, Dell, HPE, Lenovo, SpaceX, and CoreWeave have been announced as partners in the initiative. In addition, Nvidia has begun cooperating with Anthropic to integrate its cloud-based AI agents with OpenShell.