NVIDIA launched Open Agent Safety, a platform for controlling AI agents that combines software isolation and hardware monitoring. The company hopes to make autonomous systems more secure while developing its own infrastructure.
The platform is based on OpenShell, an open runtime that isolates AI agents at the core level of the system. It restricts access to files, tools, and external services, does not reveal real credentials to the agent, and logs decisions about granting permissions. The second component, Sentry, runs on the BlueField-4 DPU accelerator and monitors the agent’s operations from outside the host system. If someone attempts to violate established restrictions, the device is able to quarantine them within milliseconds.
The peculiarity of Sentry is that it requires access to the reasoning chain of the model. This may be easier to achieve using open models, whereas Anthropic and OpenAI restrict access to such materials. At the same time, NVIDIA is promoting the use of open weight models that companies can use their own computing power to run and monitor alongside the data.
The new system also connects AI security to NVIDIA hardware solutions: Sentry requires BlueField-4 to operate. Company head Jensen Huang called the refinement of the model competition, not theft, and noted that competition helps technology develop. As a result, NVIDIA is simultaneously providing a new way to protect AI agents and expanding the scope of its own infrastructure.