Nvidia unveils platform to stop AI agents from acting beyond authority

Nvidia announced a new platform designed to prevent AI agents from carrying out unauthorized actions. The announcement came during the GTC Berlin conference held in Germany on Wednesday 21 October 2026. The platform, called OpenShell Reference Architecture, allows customers to develop, run and monitor AI agents while imposing security controls that block them from exceeding assigned permissions. Nvidia said the design provides a technical foundation that companies can adapt for servers or AI‑dependent systems. The system runs AI models on the Vera Rubin platform, executes agent software on Vera chips and uses BlueField‑4 data processing units to monitor operations and stop any unauthorized activity.

Nvidia is collaborating with several firms to deliver the technology, including CoreWeave, Crusoe, Lambda, Nebius and Oracle Cloud Infrastructure. CEO Jensen Huang stated that advancing AI capabilities must be matched by stronger safety measures as institutions prepare to run billions of agents. He added that OpenShell gives agents the ability to perform tasks within clear limits while ensuring they remain monitored and controllable. The new design enables agents to process information with Rubin, execute tasks through Vera and apply protection controls through BlueField. Nvidia had previously announced OpenShell and Nvidia Sentry in late September, noting that these solutions could have prevented earlier security incidents such as a breach involving an OpenAI agent targeting Hugging Face.