Nvidia Corp. introduced the Nvidia Open Agent Safety Platform early today, a free, open-source product that packages tools for securing AI agents, tightening control over them and improving governance. The platform is aimed at organizations deploying autonomous software and arrives as high-profile incidents involving uncontrolled agents draw scrutiny from governments and enterprises alike.
The platform combines two main pieces. It is built on Nvidia's open-source OpenShell software, a runtime environment that runs fleets of autonomous agents inside isolated sandboxes and applies infrastructure-level controls. Nvidia says OpenShell lets users trace agent actions and enforce security policies outside the guardrails attached to the underlying large language model. OpenShell has been optimized for Nvidia's Vera central processing units and also works with CPUs from Intel Corp. and Arm Ltd.
The other component, Nvidia Sentry, is a reference system design based on Nvidia DOCA, the data center infrastructure-on-a-chip architecture used to write applications for Nvidia's BlueField data processing units, much as CUDA serves graphics processing units. Running on BlueField-4, Sentry monitors agents in the background, inspects requests, verifies identities and actions, and can shut agents down immediately if they cross assigned boundaries. Nvidia argues that recent incidents show traditional application-layer guardrails are easy for agents to circumvent, and that controls are needed across the agent, compute and hardware layers.
Nvidia founder and Chief Executive Jensen Huang said safety and security require full-stack engineering. He said the platform brings together industry, researchers and public-sector organizations to share best practices, align on evaluation methods and foster international cooperation, raising the bar for global AI safety.
Adoption already spans several named customers. SpaceXAI Corp. has used the platform to secure Cursor's coding agents and the Grok large language model; President Mike Nicolls said safety should be enforced outside the model through controls an agent cannot get past, and that customers should be able to set limits for Cursor and Grok and trust they will hold. Salesforce has integrated OpenShell with Slack to provide human approval controls, visibility and audit tracking, and SAP SE is using the software with the SAP Business AI Platform for runtime security on its autonomous agents.
Robotics companies Figure AI Inc., Gecko Robotics Inc. and Skild AI Inc. are relying on OpenShell to embed safety controls into systems powering autonomous robots in the real world, according to Nvidia. The platform is available for download through Nvidia's developer resources and GitHub. It remains to be seen how reliable it proves at large scale, given the ingenuity some agents have demonstrated.
More AI news from TechManNews.
