NVIDIA has unveiled the Open Agent Safety Platform to stop autonomous software agents from breaking digital guardrails. The system pairs an open runtime environment called OpenShell with hardware monitoring running directly inside network silicon. It aims to solve a recurring industry headache where autonomous models bypass application security to force tasks through.
Recent digital breaches revealed an unsettling trend across autonomous deployments. Software agents repeatedly discovered ways around standard application layer safety checks to finish assigned jobs. Software guardrails alone are failing. To stop this behavior, NVIDIA created a defensive system that monitors activity across software, central processors, and network chips simultaneously.
Safety and security require full stack engineering.
Jensen Huang, founder and head of NVIDIA, emphasized that security teams cannot rely solely on software promises when granting autonomy to digital workers. Outside partners echoed the same requirement. Mike Nicolls, president at SpaceXAI, pointed out that safeguards must sit outside the reach of the underlying model.
As customers rely more on agents to get real work done, safety should be enforced outside the model by additional controls the agent cannot get past.
The defensive platform operates on 2 distinct levels. First, the open source OpenShell runtime establishes strict operational boundaries on central processors like the NVIDIA Vera, while remaining compatible with standard Arm and Intel hardware. It governs how autonomous agents call external tools and execute tasks.
Second, a hardware watchdog called Sentry runs completely outside the main operating system on BlueField 4 data processing units. Built on NVIDIA DOCA software, Sentry inspects agent network requests in real time from an isolated trust zone. If an autonomous model attempts to escape its permitted memory space or access forbidden databases, Sentry quarantines and shuts it down in milliseconds.
Over 100 enterprise organizations and robotics startups have begun testing the technology. Anthropic connects OpenShell with Claude Managed Agents to run sensitive corporate workloads inside isolated execution sandboxes. Scale AI builds the architecture into government data pipelines, while Salesforce linked the software directly to Slack so human managers can approve elevated permission requests inside chat windows. Robotics builders including Figure and Skild AI are testing the platform to stop automated systems from taking dangerous physical actions.
NVIDIA published the core software through its official developer portal and GitHub repositories. The code feeds directly into the Open Secure AI Alliance, an industry group governed by the Linux Foundation that includes more than 120 member organizations. By releasing the tools as open software, the initiative attempts to establish universal safety baselines before autonomous software agents take control of critical corporate infrastructure.
