NVIDIA is offering a way to fence in AI agents, AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions. On Monday, Sept. 28, the company announced NVIDIA Open Agent Safety Platform, an open software platform and reference system design meant to secure agents from testing through deployment. NVIDIA presents it as an answer to a run of incidents in which agents slipped past the safety controls built into their software, and it comes from the company that makes many of the chips AI runs on.
The platform has two named parts. OpenShell is open-source software, free for anyone to use, study, change and share, that sets enforceable limits on agents running on a computer's main processors. It is tuned for NVIDIA's Vera processor and can be extended to Arm and Intel systems. Sentry is a watchdog that runs on NVIDIA's BlueField-4 data processing unit, separate from the systems it watches, where it monitors agent behavior and, NVIDIA says, can isolate a runaway agent in milliseconds.
NVIDIA says OpenShell is broadly available now. Sentry comes as a reference system design, a blueprint for building the watchdog into machines. NVIDIA says more than 100 organizations are working with the platform's technologies, and that SpaceXAI is using it for Cursor coding agents and Grok models.
Jensen Huang, NVIDIA's founder and CEO, said AI's potential "will only be realized if we solve AI safety" and that "safety and security require full-stack engineering," meaning protection designed into every layer from the chips to the software.
It is not yet clear how widely Sentry will be built beyond NVIDIA's reference design, which customers will switch the watchdog on in their own BlueField-4 systems, or whether outside labs will use the platform to contain agents during safety testing.