OpenAI is working with Nvidia on OpenShell, an open-source sandbox intended to stop AI agents from escaping, despite not joining the more than 100 organizations publicly supporting Nvidia’s broader Open Agent Safety Platform. Amazon, Google and Apple are also absent from the supporter list.
OpenShell isolates agent activity in software and can be adapted to hardware from other vendors. Hugging Face has contributed detection for agents that are allowed to visit a site but use that access to coordinate unauthorized behavior, such as leaving messages for one another in a code repository.
Nvidia’s full platform adds Sentry, a monitoring layer running on proprietary BlueField-4 data-processing units. Because the watcher operates outside the agent’s environment, Nvidia says a model cannot tell that it is being observed and the system can shut it down quickly. That hardware dependency means the complete package is not purely open source and will run most directly on Nvidia infrastructure.
OpenAI says it supports the work and collaborates privately, while developing its own safeguards and a separate information-sharing consortium called Defense Factory. Its absence from Nvidia’s pledge therefore does not mean it rejected sandboxing. It does leave customers with several overlapping security efforts and no single industry standard for proving that containment works across models, clouds and chips.